use a thread to capture frames from the subprocess so it can be killed properly

* Add object size to the bounding box

Remove script from Dockerfile

Fix framerate command

Move default value for framerate

update dockerfile

dockerfile changes

Add person_area label to surrounding box


Update dockerfile


ffmpeg config bug


Add `person_area` label to `best_person` frame


Resolve debug view showing area label for non-persons


Add object size to the bounding box


Add object size to the bounding box

* Move object area outside of conditional to work with all object types

.dockerignore

@@ -1 +1,6 @@
-README.md
+README.md
+diagram.png
+.gitignore
+debug
+config/
+*.pyc

.github/FUNDING.yml

@@ -0,0 +1 @@
+ko_fi: blakeblackshear

Dockerfile

@@ -1,68 +1,70 @@
-FROM ubuntu:16.04
+FROM ubuntu:18.04
 
-# Install system packages
-RUN apt-get -qq update && apt-get -qq install --no-install-recommends -y python3 \ 
- python3-dev \
- python-pil \
- python-lxml \
- python-tk \
+ARG DEVICE
+
+# Install packages for apt repo
+RUN apt-get -qq update && apt-get -qq install --no-install-recommends -y \
+    apt-transport-https \
+    ca-certificates \
+    curl \
+    wget \
+    gnupg-agent \
+    dirmngr \
+    software-properties-common \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY scripts/install_odroid_repo.sh .
+
+RUN if [ "$DEVICE" = "odroid" ]; then \
+      sh /install_odroid_repo.sh; \
+    fi
+
+RUN apt-get -qq update && apt-get -qq install --no-install-recommends -y \
+ python3 \ 
+ # OpenCV dependencies
+ ffmpeg \
  build-essential \
- cmake \ 
- git \ 
- libgtk2.0-dev \ 
- pkg-config \ 
- libavcodec-dev \ 
- libavformat-dev \ 
- libswscale-dev \ 
- libtbb2 \
- libtbb-dev \ 
+ cmake \
+ unzip \
+ pkg-config \
  libjpeg-dev \
  libpng-dev \
  libtiff-dev \
- libjasper-dev \
- libdc1394-22-dev \
- x11-apps \
- wget \
- vim \
- ffmpeg \
- unzip \
+ libavcodec-dev \
+ libavformat-dev \
+ libswscale-dev \
+ libv4l-dev \
+ libxvidcore-dev \
+ libx264-dev \
+ libgtk-3-dev \
+ libatlas-base-dev \
+ gfortran \
+ python3-dev \
+ # Coral USB Python API Dependencies
+ libusb-1.0-0 \
+ python3-pip \
+ python3-pil \
+ python3-numpy \
+ libc++1 \
+ libc++abi1 \
+ libunwind8 \
+ libgcc1 \
+ # VAAPI drivers for Intel hardware accel
+ libva-drm2 libva2 i965-va-driver vainfo \
  && rm -rf /var/lib/apt/lists/* 
 
 # Install core packages 
 RUN wget -q -O /tmp/get-pip.py --no-check-certificate https://bootstrap.pypa.io/get-pip.py && python3 /tmp/get-pip.py
 RUN  pip install -U pip \
  numpy \
- matplotlib \
- notebook \
- jupyter \
- pandas \
- moviepy \
- tensorflow \
- keras \
- autovizwidget \
  Flask \
- imutils \
- paho-mqtt
-
-# Install tensorflow models object detection
-RUN GIT_SSL_NO_VERIFY=true git clone -q https://github.com/tensorflow/models /usr/local/lib/python3.5/dist-packages/tensorflow/models
-RUN wget -q -P /usr/local/src/ --no-check-certificate https://github.com/google/protobuf/releases/download/v3.5.1/protobuf-python-3.5.1.tar.gz
-
-# Download & build protobuf-python
-RUN cd /usr/local/src/ \
- && tar xf protobuf-python-3.5.1.tar.gz \
- && rm protobuf-python-3.5.1.tar.gz \
- && cd /usr/local/src/protobuf-3.5.1/ \
- && ./configure \
- && make \
- && make install \
- && ldconfig \
- && rm -rf /usr/local/src/protobuf-3.5.1/
-
-# Add dataframe display widget
-RUN jupyter nbextension enable --py --sys-prefix widgetsnbextension
+ paho-mqtt \
+ PyYAML
 
 # Download & build OpenCV
+# TODO: use multistage build to reduce image size: 
+#   https://medium.com/@denismakogon/pain-and-gain-running-opencv-application-with-golang-and-docker-on-alpine-3-7-435aa11c7aec
+#   https://www.merixstudio.com/blog/docker-multi-stage-builds-python-development/
 RUN wget -q -P /usr/local/src/ --no-check-certificate https://github.com/opencv/opencv/archive/4.0.1.zip
 RUN cd /usr/local/src/ \
  && unzip 4.0.1.zip \
@@ -73,18 +75,35 @@ RUN cd /usr/local/src/ \
  && cmake -D CMAKE_INSTALL_TYPE=Release -D CMAKE_INSTALL_PREFIX=/usr/local/ .. \
  && make -j4 \
  && make install \
+ && ldconfig \
  && rm -rf /usr/local/src/opencv-4.0.1
 
+# Download and install EdgeTPU libraries for Coral
+RUN wget https://dl.google.com/coral/edgetpu_api/edgetpu_api_latest.tar.gz -O edgetpu_api.tar.gz --trust-server-names \
+  && tar xzf edgetpu_api.tar.gz
+
+COPY scripts/install_edgetpu_api.sh edgetpu_api/install.sh
+
+RUN cd edgetpu_api \
+  && /bin/bash install.sh
+
+# Copy a python 3.6 version
+RUN cd /usr/local/lib/python3.6/dist-packages/edgetpu/swig/ \
+  && ln -s _edgetpu_cpp_wrapper.cpython-35m-arm-linux-gnueabihf.so _edgetpu_cpp_wrapper.cpython-36m-arm-linux-gnueabihf.so
+
+# symlink the model and labels
+RUN wget https://dl.google.com/coral/canned_models/mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite -O mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite --trust-server-names
+RUN wget https://dl.google.com/coral/canned_models/coco_labels.txt -O coco_labels.txt --trust-server-names
+RUN ln -s mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite /frozen_inference_graph.pb
+RUN ln -s /coco_labels.txt /label_map.pbtext
+
 # Minimize image size 
 RUN (apt-get autoremove -y; \
      apt-get autoclean -y)
 
-# Set TF object detection available
-ENV PYTHONPATH "$PYTHONPATH:/usr/local/lib/python3.5/dist-packages/tensorflow/models/research:/usr/local/lib/python3.5/dist-packages/tensorflow/models/research/slim"
-RUN cd /usr/local/lib/python3.5/dist-packages/tensorflow/models/research && protoc object_detection/protos/*.proto --python_out=.
-
 WORKDIR /opt/frigate/
 ADD frigate frigate/
 COPY detect_objects.py .
+COPY benchmark.py .
 
-CMD ["python3", "-u", "detect_objects.py"]
+CMD ["python3", "-u", "detect_objects.py"]

README.md

@@ -1,18 +1,20 @@
+<a href='https://ko-fi.com/P5P7XGO9' target='_blank'><img height='36' style='border:0px;height:36px;' src='https://az743702.vo.msecnd.net/cdn/kofi4.png?v=2' border='0' alt='Buy Me a Coffee at ko-fi.com' /></a>
+
 # Frigate - Realtime Object Detection for RTSP Cameras
+**Note:** This version requires the use of a [Google Coral USB Accelerator](https://coral.withgoogle.com/products/accelerator/)
+
 Uses OpenCV and Tensorflow to perform realtime object detection locally for RTSP cameras. Designed for integration with HomeAssistant or others via MQTT.
 
 - Leverages multiprocessing and threads heavily with an emphasis on realtime over processing every frame
-- Allows you to define specific regions (squares) in the image to look for motion/objects
-- Motion detection runs in a separate process per region and signals to object detection to avoid wasting CPU cycles looking for objects when there is no motion
-- Object detection with Tensorflow runs in a separate process per region
-- Detected objects are placed on a shared mp.Queue and aggregated into a list of recently detected objects in a separate thread
-- A person score is calculated as the sum of all scores/5
-- Motion and object info is published over MQTT for integration into HomeAssistant or others
+- Allows you to define specific regions (squares) in the image to look for objects
+- No motion detection (for now)
+- Object detection with Tensorflow runs in a separate thread
+- Object info is published over MQTT for integration into HomeAssistant as a binary sensor
 - An endpoint is available to view an MJPEG stream for debugging
 
 ![Diagram](diagram.png)
 
-## Example video
+## Example video (from older version)
 You see multiple bounding boxes because it draws bounding boxes from all frames in the past 1 second where a person was detected. Not all of the bounding boxes were from the current frame.
 [![](http://img.youtube.com/vi/nqHbCtyo4dY/0.jpg)](http://www.youtube.com/watch?v=nqHbCtyo4dY "Frigate")
 
@@ -22,24 +24,16 @@ Build the container with
 docker build -t frigate .
 ```
 
-Download a model from the [zoo](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md).
-
-Download the cooresponding label map from [here](https://github.com/tensorflow/models/tree/master/research/object_detection/data).
+The `mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite` model is included and used by default. You can use your own model and labels by mounting files in the container at `/frozen_inference_graph.pb` and `/label_map.pbtext`. Models must be compatible with the Coral according to [this](https://coral.withgoogle.com/models/).
 
 Run the container with
 ```
 docker run --rm \
--v <path_to_frozen_detection_graph.pb>:/frozen_inference_graph.pb:ro \
--v <path_to_labelmap.pbtext>:/label_map.pbtext:ro \
+--privileged \
+-v /dev/bus/usb:/dev/bus/usb \
 -v <path_to_config_dir>:/config:ro \
 -p 5000:5000 \
--e RTSP_URL='<rtsp_url>' \
--e REGIONS='<box_size_1>,<x_offset_1>,<y_offset_1>,<min_person_size_1>,<min_motion_size_1>,<mask_file_1>:<box_size_2>,<x_offset_2>,<y_offset_2>,<min_person_size_2>,<min_motion_size_2>,<mask_file_2>' \
--e MQTT_HOST='your.mqtthost.com' \
--e MQTT_USER='username' \
--e MQTT_PASS='password' \
--e MQTT_TOPIC_PREFIX='cameras/1' \
--e DEBUG='0' \
+-e RTSP_PASSWORD='password' \
 frigate:latest
 ```
 
@@ -48,100 +42,59 @@ Example docker-compose:
   frigate:
     container_name: frigate
     restart: unless-stopped
+    privileged: true
     image: frigate:latest
     volumes:
-      - <path_to_frozen_detection_graph.pb>:/frozen_inference_graph.pb:ro
-      - <path_to_labelmap.pbtext>:/label_map.pbtext:ro
+      - /dev/bus/usb:/dev/bus/usb
       - <path_to_config>:/config
     ports:
-      - "127.0.0.1:5000:5000"
+      - "5000:5000"
     environment:
-      RTSP_URL: "<rtsp_url>"
-      REGIONS: "<box_size_1>,<x_offset_1>,<y_offset_1>,<min_person_size_1>,<min_motion_size_1>,<mask_file_1>:<box_size_2>,<x_offset_2>,<y_offset_2>,<min_person_size_2>,<min_motion_size_2>,<mask_file_2>"
-      MQTT_HOST: "your.mqtthost.com"
-      MQTT_USER: "username" #optional
-      MQTT_PASS: "password" #optional
-      MQTT_TOPIC_PREFIX: "cameras/1"
-      DEBUG: "0"
+      RTSP_PASSWORD: "password"
 ```
 
-Here is an example `REGIONS` env variable:
-`350,0,300,5000,200,mask-0-300.bmp:400,350,250,2000,200,mask-350-250.bmp:400,750,250,2000,200,mask-750-250.bmp`
+A `config.yml` file must exist in the `config` directory. See example [here](config/config.yml).
 
-First region broken down (all are required):
-- `350` - size of the square (350px by 350px)
-- `0` - x coordinate of upper left corner (top left of image is 0,0)
-- `300` - y coordinate of upper left corner (top left of image is 0,0)
-- `5000` - minimum person bounding box size (width*height for bounding box of identified person)
-- `200` - minimum number of changed pixels to trigger motion
-- `mask-0-300.bmp` - a bmp file with the masked regions as pure black, must be the same size as the region
-
-Mask files go in the `/config` directory.
-
-Access the mjpeg stream at http://localhost:5000
+Access the mjpeg stream at `http://localhost:5000/<camera_name>` and the best person snapshot at `http://localhost:5000/<camera_name>/best_person.jpg`
 
 ## Integration with HomeAssistant
 ```
 camera:
   - name: Camera Last Person
     platform: generic
-    still_image_url: http://<ip>:5000/best_person.jpg
+    still_image_url: http://<ip>:5000/<camera_name>/best_person.jpg
 
 binary_sensor:
-  - name: Camera Motion
+  - name: Camera Person
     platform: mqtt
-    state_topic: "cameras/1/motion"
-    device_class: motion
-    availability_topic: "cameras/1/available"
-
-sensor:
-  - name: Camera Person Score
-    platform: mqtt
-    state_topic: "cameras/1/objects"
+    state_topic: "frigate/<camera_name>/objects"
     value_template: '{{ value_json.person }}'
-    unit_of_measurement: '%'
-    availability_topic: "cameras/1/available"
+    device_class: motion
+    availability_topic: "frigate/available"
 ```
 
 ## Tips
 - Lower the framerate of the RTSP feed on the camera to reduce the CPU usage for capturing the feed
-- Use SSDLite models to reduce CPU usage
 
 ## Future improvements
-- [ ] Build tensorflow from source for CPU optimizations
+- [x] Remove motion detection for now
+- [x] Try running object detection in a thread rather than a process
+- [x] Implement min person size again
+- [x] Switch to a config file
+- [x] Handle multiple cameras in the same container
+- [ ] Attempt to figure out coral symlinking
+- [ ] Add object list to config with min scores for mqtt
+- [ ] Move mjpeg encoding to a separate process
+- [ ] Simplify motion detection (check entire image against mask, resize instead of gaussian blur)
+- [ ] See if motion detection is even worth running
+- [ ] Scan for people across entire image rather than specfic regions
+- [ ] Dynamically resize detection area and follow people
 - [ ] Add ability to turn detection on and off via MQTT
-- [ ] MQTT motion occasionally gets stuck ON
 - [ ] Output movie clips of people for notifications, etc.
 - [ ] Integrate with homeassistant push camera
 - [ ] Merge bounding boxes that span multiple regions
-- [ ] Switch to a config file
-- [ ] Allow motion regions to be different than object detection regions
 - [ ] Implement mode to save labeled objects for training
 - [ ] Try and reduce CPU usage by simplifying the tensorflow model to just include the objects we care about
 - [ ] Look into GPU accelerated decoding of RTSP stream
 - [ ] Send video over a socket and use JSMPEG
-- [ ] Look into neural compute stick
-
-## Building Tensorflow from source for CPU optimizations
-https://www.tensorflow.org/install/source#docker_linux_builds
-used `tensorflow/tensorflow:1.12.0-devel-py3`
-
-## Optimizing the graph (cant say I saw much difference in CPU usage)
-https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/graph_transforms/README.md#optimizing-for-deployment
-```
-docker run -it -v ${PWD}:/lab -v ${PWD}/../back_camera_model/models/ssd_mobilenet_v2_coco_2018_03_29/frozen_inference_graph.pb:/frozen_inference_graph.pb:ro tensorflow/tensorflow:1.12.0-devel-py3 bash
-
-bazel build tensorflow/tools/graph_transforms:transform_graph
-
-bazel-bin/tensorflow/tools/graph_transforms/transform_graph \
---in_graph=/frozen_inference_graph.pb \
---out_graph=/lab/optimized_inception_graph.pb \
---inputs='image_tensor' \
---outputs='num_detections,detection_scores,detection_boxes,detection_classes' \
---transforms='
-  strip_unused_nodes(type=float, shape="1,300,300,3")
-  remove_nodes(op=Identity, op=CheckNumerics)
-  fold_constants(ignore_errors=true)
-  fold_batch_norms
-  fold_old_batch_norms'
-```
+- [x] Look into neural compute stick

benchmark.py

@@ -0,0 +1,20 @@
+import statistics
+import numpy as np
+from edgetpu.detection.engine import DetectionEngine
+
+# Path to frozen detection graph. This is the actual model that is used for the object detection.
+PATH_TO_CKPT = '/frozen_inference_graph.pb'
+
+# Load the edgetpu engine and labels
+engine = DetectionEngine(PATH_TO_CKPT)
+
+frame = np.zeros((300,300,3), np.uint8)
+flattened_frame = np.expand_dims(frame, axis=0).flatten()
+
+detection_times = []
+
+for x in range(0, 1000):
+    objects = engine.DetectWithInputTensor(flattened_frame, threshold=0.1, top_k=3)
+    detection_times.append(engine.get_inference_time())
+
+print("Average inference time: " + str(statistics.mean(detection_times)))

config/config.yml

@@ -0,0 +1,65 @@
+web_port: 5000
+
+mqtt:
+  host: mqtt.server.com
+  topic_prefix: frigate
+#  user: username # Optional -- Uncomment for use
+#  password: password # Optional -- Uncomment for use
+
+cameras:
+  back:
+    rtsp:
+      user: viewer
+      host: 10.0.10.10
+      port: 554
+      # values that begin with a "$" will be replaced with environment variable
+      password: $RTSP_PASSWORD
+      path: /cam/realmonitor?channel=1&subtype=2
+
+    ################
+    ## Optional mask. Must be the same dimensions as your video feed.
+    ## The mask works by looking at the bottom center of the bounding box for the detected
+    ## person in the image. If that pixel in the mask is a black pixel, it ignores it as a
+    ## false positive. In my mask, the grass and driveway visible from my backdoor camera 
+    ## are white. The garage doors, sky, and trees (anywhere it would be impossible for a 
+    ## person to stand) are black.
+    ################
+    # mask: back-mask.bmp
+
+    ################
+    # Allows you to limit the framerate within frigate for cameras that do not support
+    # custom framerates. A value of 1 tells frigate to look at every frame, 2 every 2nd frame, 
+    # 3 every 3rd frame, etc.
+    ################
+    take_frame: 1
+
+    ################
+    # Optional hardware acceleration parameters for ffmpeg. If your hardware supports it, it can
+    # greatly reduce the CPU power used to decode the video stream. You will need to determine which
+    # parameters work for your specific hardware. These may work for those with Intel hardware that
+    # supports QuickSync.
+    ################
+    # ffmpeg_hwaccel_args:
+    #   - -hwaccel
+    #   - vaapi
+    #   - -hwaccel_device
+    #   - /dev/dri/renderD128
+    #   - -hwaccel_output_format
+    #   - yuv420p
+    
+    regions:
+      - size: 350
+        x_offset: 0
+        y_offset: 300
+        min_person_area: 5000
+        threshold: 0.5
+      - size: 400
+        x_offset: 350
+        y_offset: 250
+        min_person_area: 2000
+        threshold: 0.5
+      - size: 400
+        x_offset: 750
+        y_offset: 250
+        min_person_area: 2000
+        threshold: 0.5

detect_objects.py

@@ -1,247 +1,99 @@
-import os
 import cv2
-import imutils
 import time
-import datetime
-import ctypes
-import logging
-import multiprocessing as mp
-import threading
-import json
-from contextlib import closing
+import queue
+import yaml
 import numpy as np
-from object_detection.utils import visualization_utils as vis_util
-from flask import Flask, Response, make_response, send_file
+from flask import Flask, Response, make_response
 import paho.mqtt.client as mqtt
 
-from frigate.util import tonumpyarray
-from frigate.mqtt import MqttMotionPublisher, MqttObjectPublisher
-from frigate.objects import ObjectParser, ObjectCleaner, BestPersonFrame
-from frigate.motion import detect_motion
-from frigate.video import fetch_frames, FrameTracker
-from frigate.object_detection import detect_objects
+from frigate.video import Camera
+from frigate.object_detection import PreppedQueueProcessor
 
-RTSP_URL = os.getenv('RTSP_URL')
+with open('/config/config.yml') as f:
+    CONFIG = yaml.safe_load(f)
 
-MQTT_HOST = os.getenv('MQTT_HOST')
-MQTT_USER = os.getenv('MQTT_USER')
-MQTT_PASS = os.getenv('MQTT_PASS')
-MQTT_TOPIC_PREFIX = os.getenv('MQTT_TOPIC_PREFIX')
+MQTT_HOST = CONFIG['mqtt']['host']
+MQTT_PORT = CONFIG.get('mqtt', {}).get('port', 1883)
+MQTT_TOPIC_PREFIX = CONFIG.get('mqtt', {}).get('topic_prefix', 'frigate')
+MQTT_USER = CONFIG.get('mqtt', {}).get('user')
+MQTT_PASS = CONFIG.get('mqtt', {}).get('password')
 
-# REGIONS = "350,0,300,50:400,350,250,50:400,750,250,50"
-# REGIONS = "400,350,250,50"
-REGIONS = os.getenv('REGIONS')
-
-DEBUG = (os.getenv('DEBUG') == '1')
+WEB_PORT = CONFIG.get('web_port', 5000)
+DEBUG = (CONFIG.get('debug', '0') == '1')
 
 def main():
-    DETECTED_OBJECTS = []
-    recent_motion_frames = {}
-    # Parse selected regions
-    regions = []
-    for region_string in REGIONS.split(':'):
-        region_parts = region_string.split(',')
-        region_mask_image = cv2.imread("/config/{}".format(region_parts[5]), cv2.IMREAD_GRAYSCALE)
-        region_mask = np.where(region_mask_image==[0])
-        regions.append({
-            'size': int(region_parts[0]),
-            'x_offset': int(region_parts[1]),
-            'y_offset': int(region_parts[2]),
-            'min_person_area': int(region_parts[3]),
-            'min_object_size': int(region_parts[4]),
-            'mask': region_mask,
-            # Event for motion detection signaling
-            'motion_detected': mp.Event(),
-            # create shared array for storing 10 detected objects
-            # note: this must be a double even though the value you are storing
-            #       is a float. otherwise it stops updating the value in shared
-            #       memory. probably something to do with the size of the memory block
-            'output_array': mp.Array(ctypes.c_double, 6*10)
-        })
-    # capture a single frame and check the frame shape so the correct array
-    # size can be allocated in memory
-    video = cv2.VideoCapture(RTSP_URL)
-    ret, frame = video.read()
-    if ret:
-        frame_shape = frame.shape
-    else:
-        print("Unable to capture video stream")
-        exit(1)
-    video.release()
-        
-    # compute the flattened array length from the array shape
-    flat_array_length = frame_shape[0] * frame_shape[1] * frame_shape[2]
-    # create shared array for storing the full frame image data
-    shared_arr = mp.Array(ctypes.c_uint16, flat_array_length)
-    # create shared value for storing the frame_time
-    shared_frame_time = mp.Value('d', 0.0)
-    # Lock to control access to the frame
-    frame_lock = mp.Lock()
-    # Condition for notifying that a new frame is ready
-    frame_ready = mp.Condition()
-    # Condition for notifying that motion status changed globally
-    motion_changed = mp.Condition()
-    # Condition for notifying that objects were parsed
-    objects_parsed = mp.Condition()
-    # Queue for detected objects
-    object_queue = mp.Queue()
-
-    # shape current frame so it can be treated as an image
-    frame_arr = tonumpyarray(shared_arr).reshape(frame_shape)
-
-    # start the process to capture frames from the RTSP stream and store in a shared array
-    capture_process = mp.Process(target=fetch_frames, args=(shared_arr, 
-        shared_frame_time, frame_lock, frame_ready, frame_shape, RTSP_URL))
-    capture_process.daemon = True
-
-    # for each region, start a separate process for motion detection and object detection
-    detection_processes = []
-    motion_processes = []
-    for region in regions:
-        detection_process = mp.Process(target=detect_objects, args=(shared_arr, 
-            object_queue,
-            shared_frame_time,
-            frame_lock, frame_ready,
-            region['motion_detected'],
-            frame_shape, 
-            region['size'], region['x_offset'], region['y_offset'],
-            region['min_person_area'],
-            DEBUG))
-        detection_process.daemon = True
-        detection_processes.append(detection_process)
-
-        motion_process = mp.Process(target=detect_motion, args=(shared_arr,
-            shared_frame_time,
-            frame_lock, frame_ready,
-            region['motion_detected'],
-            motion_changed,
-            frame_shape, 
-            region['size'], region['x_offset'], region['y_offset'],
-            region['min_object_size'], region['mask'],
-            DEBUG))
-        motion_process.daemon = True
-        motion_processes.append(motion_process)
-
-    # start a thread to store recent motion frames for processing
-    frame_tracker = FrameTracker(frame_arr, shared_frame_time, frame_ready, frame_lock, 
-        recent_motion_frames, motion_changed, [region['motion_detected'] for region in regions])
-    frame_tracker.start()
-
-    # start a thread to store the highest scoring recent person frame
-    best_person_frame = BestPersonFrame(objects_parsed, recent_motion_frames, DETECTED_OBJECTS, 
-        motion_changed, [region['motion_detected'] for region in regions])
-    best_person_frame.start()
-
-    # start a thread to parse objects from the queue
-    object_parser = ObjectParser(object_queue, objects_parsed, DETECTED_OBJECTS)
-    object_parser.start()
-    # start a thread to expire objects from the detected objects list
-    object_cleaner = ObjectCleaner(objects_parsed, DETECTED_OBJECTS)
-    object_cleaner.start()
-
     # connect to mqtt and setup last will
-    def on_connect(client, userdata, flags, rc): 
+    def on_connect(client, userdata, flags, rc):
         print("On connect called")
+        if rc != 0:
+            if rc == 3:
+                print ("MQTT Server unavailable")
+            elif rc == 4:
+                print ("MQTT Bad username or password")
+            elif rc == 5:
+                print ("MQTT Not authorized")
+            else:
+                print ("Unable to connect to MQTT: Connection refused. Error code: " + str(rc))
         # publish a message to signal that the service is running
         client.publish(MQTT_TOPIC_PREFIX+'/available', 'online', retain=True)
-    client = mqtt.Client()
+    client = mqtt.Client(client_id="frigate")
     client.on_connect = on_connect
     client.will_set(MQTT_TOPIC_PREFIX+'/available', payload='offline', qos=1, retain=True)
     if not MQTT_USER is None:
         client.username_pw_set(MQTT_USER, password=MQTT_PASS)
-
-    client.connect(MQTT_HOST, 1883, 60)
+    client.connect(MQTT_HOST, MQTT_PORT, 60)
     client.loop_start()
-
-    # start a thread to publish object scores (currently only person)
-    mqtt_publisher = MqttObjectPublisher(client, MQTT_TOPIC_PREFIX, objects_parsed, DETECTED_OBJECTS)
-    mqtt_publisher.start()
-
-    # start thread to publish motion status
-    mqtt_motion_publisher = MqttMotionPublisher(client, MQTT_TOPIC_PREFIX, motion_changed,
-        [region['motion_detected'] for region in regions])
-    mqtt_motion_publisher.start()
-
-    # start the process of capturing frames
-    capture_process.start()
-    print("capture_process pid ", capture_process.pid)
-
-    # start the object detection processes
-    for detection_process in detection_processes:
-        detection_process.start()
-        print("detection_process pid ", detection_process.pid)
     
-    # start the motion detection processes
-    for motion_process in motion_processes:
-        motion_process.start()
-        print("motion_process pid ", motion_process.pid)
+    # Queue for prepped frames, max size set to (number of cameras * 5)
+    max_queue_size = len(CONFIG['cameras'].items())*5
+    prepped_frame_queue = queue.Queue(max_queue_size)
+
+    cameras = {}
+    for name, config in CONFIG['cameras'].items():
+        cameras[name] = Camera(name, config, prepped_frame_queue, client, MQTT_TOPIC_PREFIX)
+
+    prepped_queue_processor = PreppedQueueProcessor(
+        cameras,
+        prepped_frame_queue
+    )
+    prepped_queue_processor.start()
+
+    for name, camera in cameras.items():
+        camera.start()
+        print("Capture process for {}: {}".format(name, camera.get_capture_pid()))
 
     # create a flask app that encodes frames a mjpeg on demand
     app = Flask(__name__)
 
-    @app.route('/best_person.jpg')
-    def best_person():
-        frame = np.zeros(frame_shape, np.uint8) if best_person_frame.best_frame is None else best_person_frame.best_frame
-        ret, jpg = cv2.imencode('.jpg', frame)
+    @app.route('/<camera_name>/best_person.jpg')
+    def best_person(camera_name):
+        best_person_frame = cameras[camera_name].get_best_person()
+        if best_person_frame is None:
+            best_person_frame = np.zeros((720,1280,3), np.uint8)
+        ret, jpg = cv2.imencode('.jpg', best_person_frame)
         response = make_response(jpg.tobytes())
         response.headers['Content-Type'] = 'image/jpg'
         return response
 
-    @app.route('/')
-    def index():
+    @app.route('/<camera_name>')
+    def mjpeg_feed(camera_name):
         # return a multipart response
-        return Response(imagestream(),
+        return Response(imagestream(camera_name),
                         mimetype='multipart/x-mixed-replace; boundary=frame')
-    def imagestream():
+
+    def imagestream(camera_name):
         while True:
             # max out at 5 FPS
             time.sleep(0.2)
-            # make a copy of the current detected objects
-            detected_objects = DETECTED_OBJECTS.copy()
-            # lock and make a copy of the current frame
-            with frame_lock:
-                frame = frame_arr.copy()
-            # convert to RGB for drawing
-            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
-            # draw the bounding boxes on the screen
-            for obj in detected_objects:
-                vis_util.draw_bounding_box_on_image_array(frame,
-                    obj['ymin'],
-                    obj['xmin'],
-                    obj['ymax'],
-                    obj['xmax'],
-                    color='red',
-                    thickness=2,
-                    display_str_list=["{}: {}%".format(obj['name'],int(obj['score']*100))],
-                    use_normalized_coordinates=False)
-
-            for region in regions:
-                color = (255,255,255)
-                if region['motion_detected'].is_set():
-                    color = (0,255,0)
-                cv2.rectangle(frame, (region['x_offset'], region['y_offset']), 
-                    (region['x_offset']+region['size'], region['y_offset']+region['size']), 
-                    color, 2)
-
-            # convert back to BGR
-            frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
+            frame = cameras[camera_name].get_current_frame_with_objects()
             # encode the image into a jpg
             ret, jpg = cv2.imencode('.jpg', frame)
             yield (b'--frame\r\n'
                 b'Content-Type: image/jpeg\r\n\r\n' + jpg.tobytes() + b'\r\n\r\n')
 
-    app.run(host='0.0.0.0', debug=False)
+    app.run(host='0.0.0.0', port=WEB_PORT, debug=False)
 
-    capture_process.join()
-    for detection_process in detection_processes:
-        detection_process.join()
-    for motion_process in motion_processes:
-        motion_process.join()
-    frame_tracker.join()
-    best_person_frame.join()
-    object_parser.join()
-    object_cleaner.join()
-    mqtt_publisher.join()
+    camera.join()
 
 if __name__ == '__main__':
-    main()
+    main()

frigate/motion.py

@@ -1,109 +0,0 @@
-import datetime
-import numpy as np
-import cv2
-import imutils
-from . util import tonumpyarray
-
-# do the actual motion detection
-def detect_motion(shared_arr, shared_frame_time, frame_lock, frame_ready, motion_detected, motion_changed,
-                  frame_shape, region_size, region_x_offset, region_y_offset, min_motion_area, mask, debug):
-    # shape shared input array into frame for processing
-    arr = tonumpyarray(shared_arr).reshape(frame_shape)
-
-    avg_frame = None
-    avg_delta = None
-    frame_time = 0.0
-    motion_frames = 0
-    while True:
-        now = datetime.datetime.now().timestamp()
-        
-        with frame_ready:
-            # if there isnt a frame ready for processing or it is old, wait for a signal
-            if shared_frame_time.value == frame_time or (now - shared_frame_time.value) > 0.5:
-                frame_ready.wait()
-        
-        # lock and make a copy of the cropped frame
-        with frame_lock: 
-            cropped_frame = arr[region_y_offset:region_y_offset+region_size, region_x_offset:region_x_offset+region_size].copy().astype('uint8')
-            frame_time = shared_frame_time.value
-
-        # convert to grayscale
-        gray = cv2.cvtColor(cropped_frame, cv2.COLOR_BGR2GRAY)
-
-        # apply image mask to remove areas from motion detection
-        gray[mask] = [255]
-
-        # apply gaussian blur
-        gray = cv2.GaussianBlur(gray, (21, 21), 0)
-
-        if avg_frame is None:
-            avg_frame = gray.copy().astype("float")
-            continue
-        
-        # look at the delta from the avg_frame
-        frameDelta = cv2.absdiff(gray, cv2.convertScaleAbs(avg_frame))
-        
-        if avg_delta is None:
-            avg_delta = frameDelta.copy().astype("float")
-
-        # compute the average delta over the past few frames
-        # the alpha value can be modified to configure how sensitive the motion detection is.
-        # higher values mean the current frame impacts the delta a lot, and a single raindrop may
-        # register as motion, too low and a fast moving person wont be detected as motion
-        # this also assumes that a person is in the same location across more than a single frame
-        cv2.accumulateWeighted(frameDelta, avg_delta, 0.2)
-
-        # compute the threshold image for the current frame
-        current_thresh = cv2.threshold(frameDelta, 25, 255, cv2.THRESH_BINARY)[1]
-
-        # black out everything in the avg_delta where there isnt motion in the current frame
-        avg_delta_image = cv2.convertScaleAbs(avg_delta)
-        avg_delta_image[np.where(current_thresh==[0])] = [0]
-
-        # then look for deltas above the threshold, but only in areas where there is a delta
-        # in the current frame. this prevents deltas from previous frames from being included
-        thresh = cv2.threshold(avg_delta_image, 25, 255, cv2.THRESH_BINARY)[1]
- 
-        # dilate the thresholded image to fill in holes, then find contours
-        # on thresholded image
-        thresh = cv2.dilate(thresh, None, iterations=2)
-        cnts = cv2.findContours(thresh.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
-        cnts = imutils.grab_contours(cnts)
-
-        motion_found = False
-
-        # loop over the contours
-        for c in cnts:
-            # if the contour is big enough, count it as motion
-            contour_area = cv2.contourArea(c)
-            if contour_area > min_motion_area:
-                motion_found = True
-                if debug:
-                    cv2.drawContours(cropped_frame, [c], -1, (0, 255, 0), 2)
-                    x, y, w, h = cv2.boundingRect(c)
-                    cv2.putText(cropped_frame, str(contour_area), (x, y),
-		                cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 100, 0), 2)
-                else:
-                    break
-        
-        if motion_found:
-            motion_frames += 1
-            # if there have been enough consecutive motion frames, report motion
-            if motion_frames >= 3:
-                # only average in the current frame if the difference persists for at least 3 frames
-                cv2.accumulateWeighted(gray, avg_frame, 0.01)
-                motion_detected.set()
-                with motion_changed:
-                    motion_changed.notify_all()
-        else:
-            # when no motion, just keep averaging the frames together
-            cv2.accumulateWeighted(gray, avg_frame, 0.01)
-            motion_frames = 0
-            if motion_detected.is_set():
-                motion_detected.clear()
-                with motion_changed:
-                    motion_changed.notify_all()
-
-        if debug and motion_frames == 3:
-            cv2.imwrite("/lab/debug/motion-{}-{}-{}.jpg".format(region_x_offset, region_y_offset, datetime.datetime.now().timestamp()), cropped_frame)
-            cv2.imwrite("/lab/debug/avg_delta-{}-{}-{}.jpg".format(region_x_offset, region_y_offset, datetime.datetime.now().timestamp()), avg_delta_image)

frigate/mqtt.py

@@ -1,29 +1,6 @@
 import json
 import threading
 
-class MqttMotionPublisher(threading.Thread):
-    def __init__(self, client, topic_prefix, motion_changed, motion_flags):
-        threading.Thread.__init__(self)
-        self.client = client
-        self.topic_prefix = topic_prefix
-        self.motion_changed = motion_changed
-        self.motion_flags = motion_flags
-
-    def run(self):
-        last_sent_motion = ""
-        while True:
-            with self.motion_changed:
-                self.motion_changed.wait()
-            
-            # send message for motion
-            motion_status = 'OFF'
-            if any(obj.is_set() for obj in self.motion_flags):
-                motion_status = 'ON'
-
-            if last_sent_motion != motion_status:
-                last_sent_motion = motion_status
-                self.client.publish(self.topic_prefix+'/motion', motion_status, retain=False)
-
 class MqttObjectPublisher(threading.Thread):
     def __init__(self, client, topic_prefix, objects_parsed, detected_objects):
         threading.Thread.__init__(self)
@@ -43,11 +20,11 @@ class MqttObjectPublisher(threading.Thread):
             with self.objects_parsed:
                 self.objects_parsed.wait()
 
-            # add all the person scores in detected objects and 
-            # average over past 1 seconds (5fps)
+            # add all the person scores in detected objects
             detected_objects = self._detected_objects.copy()
-            avg_person_score = sum([obj['score'] for obj in detected_objects if obj['name'] == 'person'])/5
-            payload['person'] = int(avg_person_score*100)
+            person_score = sum([obj['score'] for obj in detected_objects if obj['name'] == 'person'])
+            # if the person score is more than 100, set person to ON
+            payload['person'] = 'ON' if int(person_score*100) > 100 else 'OFF'
 
             # send message for objects if different
             new_payload = json.dumps(payload, sort_keys=True)

frigate/object_detection.py

@@ -1,114 +1,112 @@
 import datetime
+import time
 import cv2
+import threading
 import numpy as np
-import tensorflow as tf
-from object_detection.utils import label_map_util
-from object_detection.utils import visualization_utils as vis_util
+from edgetpu.detection.engine import DetectionEngine
 from . util import tonumpyarray
 
-# TODO: make dynamic?
-NUM_CLASSES = 90
 # Path to frozen detection graph. This is the actual model that is used for the object detection.
 PATH_TO_CKPT = '/frozen_inference_graph.pb'
 # List of the strings that is used to add correct label for each box.
 PATH_TO_LABELS = '/label_map.pbtext'
 
-# Loading label map
-label_map = label_map_util.load_labelmap(PATH_TO_LABELS)
-categories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=NUM_CLASSES,
-                                                            use_display_name=True)
-category_index = label_map_util.create_category_index(categories)
+# Function to read labels from text files.
+def ReadLabelFile(file_path):
+    with open(file_path, 'r') as f:
+        lines = f.readlines()
+    ret = {}
+    for line in lines:
+        pair = line.strip().split(maxsplit=1)
+        ret[int(pair[0])] = pair[1].strip()
+    return ret
 
-# do the actual object detection
-def tf_detect_objects(cropped_frame, sess, detection_graph, region_size, region_x_offset, region_y_offset, debug):
-    # Expand dimensions since the model expects images to have shape: [1, None, None, 3]
-    image_np_expanded = np.expand_dims(cropped_frame, axis=0)
-    image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
+class PreppedQueueProcessor(threading.Thread):
+    def __init__(self, cameras, prepped_frame_queue):
 
-    # Each box represents a part of the image where a particular object was detected.
-    boxes = detection_graph.get_tensor_by_name('detection_boxes:0')
-
-    # Each score represent how level of confidence for each of the objects.
-    # Score is shown on the result image, together with the class label.
-    scores = detection_graph.get_tensor_by_name('detection_scores:0')
-    classes = detection_graph.get_tensor_by_name('detection_classes:0')
-    num_detections = detection_graph.get_tensor_by_name('num_detections:0')
-
-    # Actual detection.
-    (boxes, scores, classes, num_detections) = sess.run(
-        [boxes, scores, classes, num_detections],
-        feed_dict={image_tensor: image_np_expanded})
-
-    if debug:
-        if len([value for index,value in enumerate(classes[0]) if str(category_index.get(value).get('name')) == 'person' and scores[0,index] > 0.5]) > 0:
-            vis_util.visualize_boxes_and_labels_on_image_array(
-                cropped_frame,
-                np.squeeze(boxes),
-                np.squeeze(classes).astype(np.int32),
-                np.squeeze(scores),
-                category_index,
-                use_normalized_coordinates=True,
-                line_thickness=4)
-            cv2.imwrite("/lab/debug/obj-{}-{}-{}.jpg".format(region_x_offset, region_y_offset, datetime.datetime.now().timestamp()), cropped_frame)
-
-
-    # build an array of detected objects
-    objects = []
-    for index, value in enumerate(classes[0]):
-        score = scores[0, index]
-        if score > 0.5:
-            box = boxes[0, index].tolist()
-            objects.append({
-                        'name': str(category_index.get(value).get('name')),
-                        'score': float(score),
-                        'ymin': int((box[0] * region_size) + region_y_offset),
-                        'xmin': int((box[1] * region_size) + region_x_offset),
-                        'ymax': int((box[2] * region_size) + region_y_offset),
-                        'xmax': int((box[3] * region_size) + region_x_offset)
-                    })
-
-    return objects
-
-def detect_objects(shared_arr, object_queue, shared_frame_time, frame_lock, frame_ready, 
-                   motion_detected, frame_shape, region_size, region_x_offset, region_y_offset,
-                   min_person_area, debug):
-    # shape shared input array into frame for processing
-    arr = tonumpyarray(shared_arr).reshape(frame_shape)
-
-    # Load a (frozen) Tensorflow model into memory before the processing loop
-    detection_graph = tf.Graph()
-    with detection_graph.as_default():
-        od_graph_def = tf.GraphDef()
-        with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:
-            serialized_graph = fid.read()
-            od_graph_def.ParseFromString(serialized_graph)
-            tf.import_graph_def(od_graph_def, name='')
-        sess = tf.Session(graph=detection_graph)
-
-    frame_time = 0.0
-    while True:
-        now = datetime.datetime.now().timestamp()
-
-        # wait until motion is detected
-        motion_detected.wait()
-
-        with frame_ready:
-            # if there isnt a frame ready for processing or it is old, wait for a new frame
-            if shared_frame_time.value == frame_time or (now - shared_frame_time.value) > 0.5:
-                frame_ready.wait()
+        threading.Thread.__init__(self)
+        self.cameras = cameras
+        self.prepped_frame_queue = prepped_frame_queue
         
-        # make a copy of the cropped frame
-        with frame_lock:
-            cropped_frame = arr[region_y_offset:region_y_offset+region_size, region_x_offset:region_x_offset+region_size].copy()
-            frame_time = shared_frame_time.value
+        # Load the edgetpu engine and labels
+        self.engine = DetectionEngine(PATH_TO_CKPT)
+        self.labels = ReadLabelFile(PATH_TO_LABELS)
 
-        # convert to RGB
-        cropped_frame_rgb = cv2.cvtColor(cropped_frame, cv2.COLOR_BGR2RGB)
-        # do the object detection
-        objects = tf_detect_objects(cropped_frame_rgb, sess, detection_graph, region_size, region_x_offset, region_y_offset, debug)
-        for obj in objects:
-            # ignore persons below the size threshold
-            if obj['name'] == 'person' and (obj['xmax']-obj['xmin'])*(obj['ymax']-obj['ymin']) < min_person_area:
-                continue
-            obj['frame_time'] = frame_time
-            object_queue.put(obj)
+    def run(self):
+        # process queue...
+        while True:
+            frame = self.prepped_frame_queue.get()
+
+            # Actual detection.
+            objects = self.engine.DetectWithInputTensor(frame['frame'], threshold=frame['region_threshold'], top_k=3)
+            # print(self.engine.get_inference_time())
+
+            # parse and pass detected objects back to the camera
+            parsed_objects = []
+            for obj in objects:
+                box = obj.bounding_box.flatten().tolist()
+                parsed_objects.append({
+                            'frame_time': frame['frame_time'],
+                            'name': str(self.labels[obj.label_id]),
+                            'score': float(obj.score),
+                            'xmin': int((box[0] * frame['region_size']) + frame['region_x_offset']),
+                            'ymin': int((box[1] * frame['region_size']) + frame['region_y_offset']),
+                            'xmax': int((box[2] * frame['region_size']) + frame['region_x_offset']),
+                            'ymax': int((box[3] * frame['region_size']) + frame['region_y_offset'])
+                        })
+            self.cameras[frame['camera_name']].add_objects(parsed_objects)
+
+
+# should this be a region class?
+class FramePrepper(threading.Thread):
+    def __init__(self, camera_name, shared_frame, frame_time, frame_ready, 
+        frame_lock,
+        region_size, region_x_offset, region_y_offset, region_threshold,
+        prepped_frame_queue):
+
+        threading.Thread.__init__(self)
+        self.camera_name = camera_name
+        self.shared_frame = shared_frame
+        self.frame_time = frame_time
+        self.frame_ready = frame_ready
+        self.frame_lock = frame_lock
+        self.region_size = region_size
+        self.region_x_offset = region_x_offset
+        self.region_y_offset = region_y_offset
+        self.region_threshold = region_threshold
+        self.prepped_frame_queue = prepped_frame_queue
+
+    def run(self):
+        frame_time = 0.0
+        while True:
+            now = datetime.datetime.now().timestamp()
+
+            with self.frame_ready:
+                # if there isnt a frame ready for processing or it is old, wait for a new frame
+                if self.frame_time.value == frame_time or (now - self.frame_time.value) > 0.5:
+                    self.frame_ready.wait()
+            
+            # make a copy of the cropped frame
+            with self.frame_lock:
+                cropped_frame = self.shared_frame[self.region_y_offset:self.region_y_offset+self.region_size, self.region_x_offset:self.region_x_offset+self.region_size].copy()
+                frame_time = self.frame_time.value
+            
+            # Resize to 300x300 if needed
+            if cropped_frame.shape != (300, 300, 3):
+                cropped_frame = cv2.resize(cropped_frame, dsize=(300, 300), interpolation=cv2.INTER_LINEAR)
+            # Expand dimensions since the model expects images to have shape: [1, 300, 300, 3]
+            frame_expanded = np.expand_dims(cropped_frame, axis=0)
+
+            # add the frame to the queue
+            if not self.prepped_frame_queue.full():
+                self.prepped_frame_queue.put({
+                    'camera_name': self.camera_name,
+                    'frame_time': frame_time,
+                    'frame': frame_expanded.flatten().copy(),
+                    'region_size': self.region_size,
+                    'region_threshold': self.region_threshold,
+                    'region_x_offset': self.region_x_offset,
+                    'region_y_offset': self.region_y_offset
+                })
+            else:
+                print("queue full. moving on")

frigate/objects.py

@@ -2,22 +2,7 @@ import time
 import datetime
 import threading
 import cv2
-from object_detection.utils import visualization_utils as vis_util
-class ObjectParser(threading.Thread):
-    def __init__(self, object_queue, objects_parsed, detected_objects):
-        threading.Thread.__init__(self)
-        self._object_queue = object_queue
-        self._objects_parsed = objects_parsed
-        self._detected_objects = detected_objects
-
-    def run(self):
-        while True:
-            obj = self._object_queue.get()
-            self._detected_objects.append(obj)
-
-            # notify that objects were parsed
-            with self._objects_parsed:
-                self._objects_parsed.notify_all()
+from . util import draw_box_with_label
 
 class ObjectCleaner(threading.Thread):
     def __init__(self, objects_parsed, detected_objects):
@@ -28,14 +13,18 @@ class ObjectCleaner(threading.Thread):
     def run(self):
         while True:
 
+            # wait a bit before checking for expired frames
+            time.sleep(0.2)
+
             # expire the objects that are more than 1 second old
             now = datetime.datetime.now().timestamp()
             # look for the first object found within the last second
             # (newest objects are appended to the end)
             detected_objects = self._detected_objects.copy()
+
             num_to_delete = 0
             for obj in detected_objects:
-                if now-obj['frame_time']<1:
+                if now-obj['frame_time']<2:
                     break
                 num_to_delete += 1
             if num_to_delete > 0:
@@ -44,80 +33,56 @@ class ObjectCleaner(threading.Thread):
                 # notify that parsed objects were changed
                 with self._objects_parsed:
                     self._objects_parsed.notify_all()
-            
-            # wait a bit before checking for more expired frames
-            time.sleep(0.2)
+
 
 # Maintains the frame and person with the highest score from the most recent
 # motion event
 class BestPersonFrame(threading.Thread):
-    def __init__(self, objects_parsed, recent_frames, detected_objects, motion_changed, motion_regions):
+    def __init__(self, objects_parsed, recent_frames, detected_objects):
         threading.Thread.__init__(self)
         self.objects_parsed = objects_parsed
         self.recent_frames = recent_frames
         self.detected_objects = detected_objects
-        self.motion_changed = motion_changed
-        self.motion_regions = motion_regions
         self.best_person = None
         self.best_frame = None
 
     def run(self):
-        motion_start = 0.0
-        motion_end = 0.0
-
         while True:
 
-             # while there is motion
-            while len([r for r in self.motion_regions if r.is_set()]) > 0:
-                # wait until objects have been parsed
-                with self.objects_parsed:
-                    self.objects_parsed.wait()
+            # wait until objects have been parsed
+            with self.objects_parsed:
+                self.objects_parsed.wait()
 
-                # make a copy of detected objects
-                detected_objects = self.detected_objects.copy()
-                detected_people = [obj for obj in detected_objects if obj['name'] == 'person']
-                # make a copy of the recent frames
-                recent_frames = self.recent_frames.copy()
+            # make a copy of detected objects
+            detected_objects = self.detected_objects.copy()
+            detected_people = [obj for obj in detected_objects if obj['name'] == 'person']
 
-                # get the highest scoring person
-                new_best_person = max(detected_people, key=lambda x:x['score'], default=self.best_person)
+            # get the highest scoring person
+            new_best_person = max(detected_people, key=lambda x:x['score'], default=self.best_person)
 
-                # if there isnt a person, continue
-                if new_best_person is None:
-                    continue
+            # if there isnt a person, continue
+            if new_best_person is None:
+                continue
 
-                # if there is no current best_person
-                if self.best_person is None:
+            # if there is no current best_person
+            if self.best_person is None:
+                self.best_person = new_best_person
+            # if there is already a best_person
+            else:
+                now = datetime.datetime.now().timestamp()
+                # if the new best person is a higher score than the current best person 
+                # or the current person is more than 1 minute old, use the new best person
+                if new_best_person['score'] > self.best_person['score'] or (now - self.best_person['frame_time']) > 60:
                     self.best_person = new_best_person
-                # if there is already a best_person
-                else:
-                    now = datetime.datetime.now().timestamp()
-                    # if the new best person is a higher score than the current best person 
-                    # or the current person is more than 1 minute old, use the new best person
-                    if new_best_person['score'] > self.best_person['score'] or (now - self.best_person['frame_time']) > 60:
-                        self.best_person = new_best_person
-
-                if not self.best_person is None and self.best_person['frame_time'] in recent_frames:
-                    best_frame = recent_frames[self.best_person['frame_time']]
-                    best_frame = cv2.cvtColor(best_frame, cv2.COLOR_BGR2RGB)
-                    # draw the bounding box on the frame
-                    vis_util.draw_bounding_box_on_image_array(best_frame,
-                        self.best_person['ymin'],
-                        self.best_person['xmin'],
-                        self.best_person['ymax'],
-                        self.best_person['xmax'],
-                        color='red',
-                        thickness=2,
-                        display_str_list=["{}: {}%".format(self.best_person['name'],int(self.best_person['score']*100))],
-                        use_normalized_coordinates=False)
-
-                    # convert back to BGR
-                    self.best_frame = cv2.cvtColor(best_frame, cv2.COLOR_RGB2BGR)
-
-            motion_end = datetime.datetime.now().timestamp()
-
-            # wait for the global motion flag to change
-            with self.motion_changed:
-                self.motion_changed.wait()
             
-            motion_start = datetime.datetime.now().timestamp()
+            # make a copy of the recent frames
+            recent_frames = self.recent_frames.copy()
+            
+            if not self.best_person is None and self.best_person['frame_time'] in recent_frames:
+                best_frame = recent_frames[self.best_person['frame_time']]
+
+                label = "{}: {}% {}".format(self.best_person['name'],int(self.best_person['score']*100),int(self.best_person['area']))
+                draw_box_with_label(best_frame, self.best_person['xmin'], self.best_person['ymin'], 
+                    self.best_person['xmax'], self.best_person['ymax'], label)
+                
+                self.best_frame = cv2.cvtColor(best_frame, cv2.COLOR_RGB2BGR)

frigate/util.py

@@ -1,5 +1,26 @@
 import numpy as np
+import cv2
 
 # convert shared memory array into numpy array
 def tonumpyarray(mp_arr):
-    return np.frombuffer(mp_arr.get_obj(), dtype=np.uint16)
+    return np.frombuffer(mp_arr.get_obj(), dtype=np.uint8)
+
+def draw_box_with_label(frame, x_min, y_min, x_max, y_max, label):
+    color = (255,0,0)
+    cv2.rectangle(frame, (x_min, y_min), 
+        (x_max, y_max), 
+        color, 2)
+    font_scale = 0.5
+    font = cv2.FONT_HERSHEY_SIMPLEX
+    # get the width and height of the text box
+    size = cv2.getTextSize(label, font, fontScale=font_scale, thickness=2)
+    text_width = size[0][0]
+    text_height = size[0][1]
+    line_height = text_height + size[1]
+    # set the text start position
+    text_offset_x = x_min
+    text_offset_y = 0 if y_min < line_height else y_min - (line_height+8)
+    # make the coords of the box with a small padding of two pixels
+    textbox_coords = ((text_offset_x, text_offset_y), (text_offset_x + text_width + 2, text_offset_y + line_height))
+    cv2.rectangle(frame, textbox_coords[0], textbox_coords[1], color, cv2.FILLED)
+    cv2.putText(frame, label, (text_offset_x, text_offset_y + line_height - 3), font, fontScale=font_scale, color=(0, 0, 0), thickness=2)

frigate/video.py

@@ -1,95 +1,323 @@
+import os
 import time
 import datetime
 import cv2
 import threading
-from . util import tonumpyarray
-
-# fetch the frames as fast a possible, only decoding the frames when the
-# detection_process has consumed the current frame
-def fetch_frames(shared_arr, shared_frame_time, frame_lock, frame_ready, frame_shape, rtsp_url):
-    # convert shared memory array into numpy and shape into image array
-    arr = tonumpyarray(shared_arr).reshape(frame_shape)
-
-    # start the video capture
-    video = cv2.VideoCapture()
-    video.open(rtsp_url)
-    # keep the buffer small so we minimize old data
-    video.set(cv2.CAP_PROP_BUFFERSIZE,1)
-
-    bad_frame_counter = 0
-    while True:
-        # check if the video stream is still open, and reopen if needed
-        if not video.isOpened():
-            success = video.open(rtsp_url)
-            if not success:
-                time.sleep(1)
-                continue
-        # grab the frame, but dont decode it yet
-        ret = video.grab()
-        # snapshot the time the frame was grabbed
-        frame_time = datetime.datetime.now()
-        if ret:
-            # go ahead and decode the current frame
-            ret, frame = video.retrieve()
-            if ret:
-                # Lock access and update frame
-                with frame_lock:
-                    arr[:] = frame
-                    shared_frame_time.value = frame_time.timestamp()
-                # Notify with the condition that a new frame is ready
-                with frame_ready:
-                    frame_ready.notify_all()
-                bad_frame_counter = 0
-            else:
-                print("Unable to decode frame")
-                bad_frame_counter += 1
-        else:
-            print("Unable to grab a frame")
-            bad_frame_counter += 1
-        
-        if bad_frame_counter > 100:
-            video.release()
-    
-    video.release()
+import ctypes
+import multiprocessing as mp
+import subprocess as sp
+import numpy as np
+from . util import tonumpyarray, draw_box_with_label
+from . object_detection import FramePrepper
+from . objects import ObjectCleaner, BestPersonFrame
+from . mqtt import MqttObjectPublisher
 
 # Stores 2 seconds worth of frames when motion is detected so they can be used for other threads
 class FrameTracker(threading.Thread):
-    def __init__(self, shared_frame, frame_time, frame_ready, frame_lock, recent_frames, motion_changed, motion_regions):
+    def __init__(self, shared_frame, frame_time, frame_ready, frame_lock, recent_frames):
         threading.Thread.__init__(self)
         self.shared_frame = shared_frame
         self.frame_time = frame_time
         self.frame_ready = frame_ready
         self.frame_lock = frame_lock
         self.recent_frames = recent_frames
-        self.motion_changed = motion_changed
-        self.motion_regions = motion_regions
 
     def run(self):
         frame_time = 0.0
         while True:
-            # while there is motion
-            while len([r for r in self.motion_regions if r.is_set()]) > 0:
-                now = datetime.datetime.now().timestamp()
-                # wait for a frame
-                with self.frame_ready:
-                    # if there isnt a frame ready for processing or it is old, wait for a signal
-                    if self.frame_time.value == frame_time or (now - self.frame_time.value) > 0.5:
-                        self.frame_ready.wait()
-                
-                # lock and make a copy of the frame
-                with self.frame_lock: 
-                    frame = self.shared_frame.copy().astype('uint8')
-                    frame_time = self.frame_time.value
-                
-                # add the frame to recent frames
-                self.recent_frames[frame_time] = frame
+            now = datetime.datetime.now().timestamp()
+            # wait for a frame
+            with self.frame_ready:
+                # if there isnt a frame ready for processing or it is old, wait for a signal
+                if self.frame_time.value == frame_time or (now - self.frame_time.value) > 0.5:
+                    self.frame_ready.wait()
+            
+            # lock and make a copy of the frame
+            with self.frame_lock: 
+                frame = self.shared_frame.copy()
+                frame_time = self.frame_time.value
+            
+            # add the frame to recent frames
+            self.recent_frames[frame_time] = frame
 
-                # delete any old frames
-                stored_frame_times = list(self.recent_frames.keys())
-                for k in stored_frame_times:
-                    if (now - k) > 2:
-                        del self.recent_frames[k]
+            # delete any old frames
+            stored_frame_times = list(self.recent_frames.keys())
+            for k in stored_frame_times:
+                if (now - k) > 2:
+                    del self.recent_frames[k]
+
+def get_frame_shape(rtsp_url):
+    # capture a single frame and check the frame shape so the correct array
+    # size can be allocated in memory
+    video = cv2.VideoCapture(rtsp_url)
+    ret, frame = video.read()
+    frame_shape = frame.shape
+    video.release()
+    return frame_shape
+
+def get_rtsp_url(rtsp_config):
+    if (rtsp_config['password'].startswith('$')):
+        rtsp_config['password'] = os.getenv(rtsp_config['password'][1:])
+    return 'rtsp://{}:{}@{}:{}{}'.format(rtsp_config['user'], 
+        rtsp_config['password'], rtsp_config['host'], rtsp_config['port'],
+        rtsp_config['path'])
+
+class CameraWatchdog(threading.Thread):
+    def __init__(self, camera):
+        threading.Thread.__init__(self)
+        self.camera = camera
+
+    def run(self):
+
+        while True:
+            # wait a bit before checking
+            time.sleep(10)
+
+            if (datetime.datetime.now().timestamp() - self.camera.frame_time.value) > 2:
+                print("last frame is more than 2 seconds old, restarting camera capture...")
+                self.camera.start_or_restart_capture()
+                time.sleep(5)
+
+# Thread to read the stdout of the ffmpeg process and update the current frame
+class CameraCapture(threading.Thread):
+    def __init__(self, camera):
+        threading.Thread.__init__(self)
+        self.camera = camera
+
+    def run(self):
+        frame_num = 0
+        while True:
+            if self.camera.ffmpeg_process.poll() != None:
+                print("ffmpeg process is not running. exiting capture thread...")
+                break
+
+            raw_image = self.camera.ffmpeg_process.stdout.read(self.camera.frame_size)
+
+            if len(raw_image) == 0:
+                print("ffmpeg didnt return a frame. something is wrong. exiting capture thread...")
+                break
+
+            frame_num += 1
+            if (frame_num % self.camera.take_frame) != 0:
+                continue
+
+            with self.camera.frame_lock:
+                self.camera.frame_time.value = datetime.datetime.now().timestamp()
                 
-            # wait for the global motion flag to change
-            with self.motion_changed:
-                self.motion_changed.wait()
+                self.camera.current_frame[:] = (
+                    np
+                    .frombuffer(raw_image, np.uint8)
+                    .reshape(self.camera.frame_shape)
+                )
+            # Notify with the condition that a new frame is ready
+            with self.camera.frame_ready:
+                self.camera.frame_ready.notify_all()
+
+class Camera:
+    def __init__(self, name, config, prepped_frame_queue, mqtt_client, mqtt_prefix):
+        self.name = name
+        self.config = config
+        self.detected_objects = []
+        self.recent_frames = {}
+        self.rtsp_url = get_rtsp_url(self.config['rtsp'])
+        self.take_frame = self.config.get('take_frame', 1)
+        self.ffmpeg_hwaccel_args = self.config.get('ffmpeg_hwaccel_args', [])
+        self.regions = self.config['regions']
+        self.frame_shape = get_frame_shape(self.rtsp_url)
+        self.frame_size = self.frame_shape[0] * self.frame_shape[1] * self.frame_shape[2]
+        self.mqtt_client = mqtt_client
+        self.mqtt_topic_prefix = '{}/{}'.format(mqtt_prefix, self.name)
+
+        # create a numpy array for the current frame in initialize to zeros
+        self.current_frame = np.zeros(self.frame_shape, np.uint8)
+        # create shared value for storing the frame_time
+        self.frame_time = mp.Value('d', 0.0)
+        # Lock to control access to the frame
+        self.frame_lock = mp.Lock()
+        # Condition for notifying that a new frame is ready
+        self.frame_ready = mp.Condition()
+        # Condition for notifying that objects were parsed
+        self.objects_parsed = mp.Condition()
+
+        self.ffmpeg_process = None
+        self.capture_thread = None
+
+        # for each region, create a separate thread to resize the region and prep for detection
+        self.detection_prep_threads = []
+        for region in self.config['regions']:
+            # set a default threshold of 0.5 if not defined
+            if not 'threshold' in region:
+                region['threshold'] = 0.5
+            if not isinstance(region['threshold'], float):
+                print('Threshold is not a float. Setting to 0.5 default.')
+                region['threshold'] = 0.5
+            self.detection_prep_threads.append(FramePrepper(
+                self.name,
+                self.current_frame,
+                self.frame_time,
+                self.frame_ready,
+                self.frame_lock,
+                region['size'], region['x_offset'], region['y_offset'], region['threshold'],
+                prepped_frame_queue
+            ))
+        
+        # start a thread to store recent motion frames for processing
+        self.frame_tracker = FrameTracker(self.current_frame, self.frame_time, 
+            self.frame_ready, self.frame_lock, self.recent_frames)
+        self.frame_tracker.start()
+
+        # start a thread to store the highest scoring recent person frame
+        self.best_person_frame = BestPersonFrame(self.objects_parsed, self.recent_frames, self.detected_objects)
+        self.best_person_frame.start()
+
+        # start a thread to expire objects from the detected objects list
+        self.object_cleaner = ObjectCleaner(self.objects_parsed, self.detected_objects)
+        self.object_cleaner.start()
+
+        # start a thread to publish object scores (currently only person)
+        mqtt_publisher = MqttObjectPublisher(self.mqtt_client, self.mqtt_topic_prefix, self.objects_parsed, self.detected_objects)
+        mqtt_publisher.start()
+
+        # create a watchdog thread for capture process
+        self.watchdog = CameraWatchdog(self)
+
+        # load in the mask for person detection
+        if 'mask' in self.config:
+            self.mask = cv2.imread("/config/{}".format(self.config['mask']), cv2.IMREAD_GRAYSCALE)
+        else:
+            self.mask = None
+
+        if self.mask is None:
+            self.mask = np.zeros((self.frame_shape[0], self.frame_shape[1], 1), np.uint8)
+            self.mask[:] = 255
+
+
+    def start_or_restart_capture(self):
+        if not self.ffmpeg_process is None:
+            print("Killing the existing ffmpeg process...")
+            self.ffmpeg_process.kill()
+            self.ffmpeg_process.wait()
+            print("Waiting for the capture thread to exit...")
+            self.capture_thread.join()
+            self.ffmpeg_process = None
+            self.capture_thread = None
+            
+        # create the process to capture frames from the RTSP stream and store in a shared array
+        print("Creating a new ffmpeg process...")
+        self.start_ffmpeg()
+        
+        print("Creating a new capture thread...")
+        self.capture_thread = CameraCapture(self)
+        print("Starting a new capture thread...")
+        self.capture_thread.start()
+    
+    def start_ffmpeg(self):
+        ffmpeg_global_args = [
+            '-hide_banner', '-loglevel', 'panic'
+        ]
+        ffmpeg_input_args = [
+            '-avoid_negative_ts', 'make_zero', 
+            '-fflags', 'nobuffer',
+            '-flags', 'low_delay',
+            '-strict', 'experimental',
+            '-fflags', '+genpts', 
+            '-rtsp_transport', 'tcp', 
+            '-stimeout', '5000000', 
+            '-use_wallclock_as_timestamps', '1'
+        ]
+
+        ffmpeg_cmd = (['ffmpeg'] +
+            ffmpeg_global_args +
+            self.ffmpeg_hwaccel_args +
+            ffmpeg_input_args +
+            ['-i', self.rtsp_url,
+            '-f', 'rawvideo',
+            '-pix_fmt', 'rgb24',
+            'pipe:'])
+
+        print(" ".join(ffmpeg_cmd))
+        
+        self.ffmpeg_process = sp.Popen(ffmpeg_cmd, stdout = sp.PIPE, bufsize=self.frame_size)
+    
+    def start(self):
+        self.start_or_restart_capture()
+        # start the object detection prep threads
+        for detection_prep_thread in self.detection_prep_threads:
+            detection_prep_thread.start()
+        self.watchdog.start()
+    
+    def join(self):
+        self.capture_thread.join()
+    
+    def get_capture_pid(self):
+        return self.ffmpeg_process.pid
+    
+    def add_objects(self, objects):
+        if len(objects) == 0:
+            return
+
+        for obj in objects:
+            # Store object area to use in bounding box labels
+            obj['area'] = (obj['xmax']-obj['xmin'])*(obj['ymax']-obj['ymin'])
+
+            if obj['name'] == 'person':
+                # find the matching region
+                region = None
+                for r in self.regions:
+                    if (
+                            obj['xmin'] >= r['x_offset'] and
+                            obj['ymin'] >= r['y_offset'] and
+                            obj['xmax'] <= r['x_offset']+r['size'] and
+                            obj['ymax'] <= r['y_offset']+r['size']
+                        ): 
+                        region = r
+                        break
+                
+                # if the min person area is larger than the
+                # detected person, don't add it to detected objects
+                if region and 'min_person_area' in region and region['min_person_area'] > obj['area']:
+                    continue
+            
+                # compute the coordinates of the person and make sure
+                # the location isnt outside the bounds of the image (can happen from rounding)
+                y_location = min(int(obj['ymax']), len(self.mask)-1)
+                x_location = min(int((obj['xmax']-obj['xmin'])/2.0)+obj['xmin'], len(self.mask[0])-1)
+
+                # if the person is in a masked location, continue
+                if self.mask[y_location][x_location] == [0]:
+                    continue
+
+            self.detected_objects.append(obj)
+
+        with self.objects_parsed:
+            self.objects_parsed.notify_all()
+
+    def get_best_person(self):
+        return self.best_person_frame.best_frame
+    
+    def get_current_frame_with_objects(self):
+        # make a copy of the current detected objects
+        detected_objects = self.detected_objects.copy()
+        # lock and make a copy of the current frame
+        with self.frame_lock:
+            frame = self.current_frame.copy()
+
+        # draw the bounding boxes on the screen
+        for obj in detected_objects:
+            label = "{}: {}% {}".format(obj['name'],int(obj['score']*100),int(obj['area']))
+            draw_box_with_label(frame, obj['xmin'], obj['ymin'], obj['xmax'], obj['ymax'], label)
+
+        for region in self.regions:
+            color = (255,255,255)
+            cv2.rectangle(frame, (region['x_offset'], region['y_offset']), 
+                (region['x_offset']+region['size'], region['y_offset']+region['size']), 
+                color, 2)
+
+        # convert to BGR
+        frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
+
+        return frame
+
+
+    
+        

scripts/install_edgetpu_api.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+set -e
+
+CPU_ARCH=$(uname -m)
+OS_VERSION=$(uname -v)
+
+echo "CPU_ARCH ${CPU_ARCH}"
+echo "OS_VERSION ${OS_VERSION}"
+
+if [[ "${CPU_ARCH}" == "x86_64" ]]; then
+  echo "Recognized as Linux on x86_64."
+  LIBEDGETPU_SUFFIX=x86_64
+  HOST_GNU_TYPE=x86_64-linux-gnu
+elif [[ "${CPU_ARCH}" == "armv7l" ]]; then
+  echo "Recognized as Linux on ARM32 platform."
+  LIBEDGETPU_SUFFIX=arm32
+  HOST_GNU_TYPE=arm-linux-gnueabihf
+elif [[ "${CPU_ARCH}" == "aarch64" ]]; then
+  echo "Recognized as generic ARM64 platform."
+  LIBEDGETPU_SUFFIX=arm64
+  HOST_GNU_TYPE=aarch64-linux-gnu
+fi
+
+if [[ -z "${HOST_GNU_TYPE}" ]]; then
+  echo "Your platform is not supported."
+  exit 1
+fi
+
+echo "Using maximum operating frequency."
+LIBEDGETPU_SRC="libedgetpu/libedgetpu_${LIBEDGETPU_SUFFIX}.so"
+LIBEDGETPU_DST="/usr/lib/${HOST_GNU_TYPE}/libedgetpu.so.1.0"
+
+# Runtime library.
+echo "Installing Edge TPU runtime library [${LIBEDGETPU_DST}]..."
+if [[ -f "${LIBEDGETPU_DST}" ]]; then
+  echo "File already exists. Replacing it..."
+  rm -f "${LIBEDGETPU_DST}"
+fi
+
+cp -p "${LIBEDGETPU_SRC}" "${LIBEDGETPU_DST}"
+ldconfig
+echo "Done."
+
+# Python API.
+WHEEL=$(ls edgetpu-*-py3-none-any.whl 2>/dev/null)
+if [[ $? == 0 ]]; then
+  echo "Installing Edge TPU Python API..."
+  python3 -m pip install --no-deps "${WHEEL}"
+  echo "Done."
+fi

scripts/install_odroid_repo.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+
+apt-key adv --keyserver keyserver.ubuntu.com --recv-keys D986B59D
+
+echo "deb http://deb.odroid.in/5422-s bionic main" > /etc/apt/sources.list.d/odroid.list