ODaM - Object Detection and Monitoring

v0.8.3

ODaM is project aimed to do monitoring such as: pedestrian detection and counting, vehicle detection and counting, speed estimation of objects, sending detected objects to gRPC server for detailed analysis.

It's written on Go with a lot of CGO.

YOLOv4 + Kalman filter for tracking	YOLOv4 + simple centroid tracking

YOLOv4 Tiny + Kalman filter for tracking	YOLOv4 Tiny + simple centroid tracking

Work in progress

We are working on this.

Not too fast, but it is what it is.

Table of Contents

• About
• QA section
• Installation
• Usage
• Screenshots
• Support
• Roadmap
• Dependencies
• License
• Devs

About

ODaM is tool for doing monitoring via Darknet's neural network called Yolo V4 (paper: https://arxiv.org/abs/2004.10934).

It's built on top of go-darknet which uses AlexeyAB's fork of Darknet. For doing computer vision stuff and video reading GoCV is used.

QA section

Who are you and what do you do?

There is info about me here: https://github.com/LdDl

You can have chat with me in Telegram/Gmail

Is this library / software or even framework?

I think about it as software with library capabilities.

What it capable of?

Not that much currently:

• Object detection via darknet: both YOLOv3 and YOLOv4 (thanks to Go bindings for it)
• Object tracking via two possible techniques: Kalman tracking (filtering) or Centroid tracking;
• Sending data to dedicated gRPC server;
• MJPEG / imshow optional visual output;
• Speed estimation based of GIS calculations (via matching pixels to WGS84).

Why Go?

Well, C++ is a killer in computer vision field and Python has a great battery included bindings for C++ code.

But I do no think that I'm ready to build gRPC/REST or any other web components of this software in C++ or Python (C++ is not that easy and Python...I just don't like Python syntax). That's why I prefer to stick with Go.

Why did you pick JSON for configuration purposes instead of TOML/YAML/INI or any other well-suited formats?

1. Compared to TOML, JSON is not that 'human friendly', but still readable.
2. It is in standart Go's library.
3. Well, it is in standart Go's library.
4. You got the idea.

Why bindings to Darknet instead of Opencv included stuff?

Sometimes you just do not need full OpenCV installation for object detection. I have such ANPR projet here: https://github.com/LdDl/license_plate_recognition I guess when I'm done with stable core I might switch from Go's Darknet bindings to OpenCV one (since ODaM-project requires OpenCV installation obviously)

What are your plans?

There is ROADMAP.md, but overall I am planning to extend capabilities of software: * Improve perfomance * Implement some cool tracking techniques (e.g. SORT) * Do gRPC accepting microservice for enabling software to catch information from external devices/systems/microservices and etc. E.g: you want to send message 'there is red light on traffic light" to instance of software, then it would look like grpcServer.Send('there is red light on traffic light'). After that any captured object will have state with message above in it. So you can catch traffic offenders. * Introduce convex polygon based calculations (same as virtual lines but for polygons)

How to help you?

If you are here, then you are already helped a lot, since you noticed my existence :harold_face:

If you want to make PR for some undone features (algorithms mainly) I'll glad to take a look.

Installation

notice: targeted for Linux users (no Windows/OSX instructions currenlty)

Need to enable CUDA (GPU) in every installation step where it's possible.

1. Install CUDA (we recommend version 10.2) bash wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-ubuntu1804.pin sudo mv cuda-ubuntu1804.pin /etc/apt/preferences.d/cuda-repository-pin-600 wget http://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda-repo-ubuntu1804-10-2-local-10.2.89-440.33.01_1.0-1_amd64.deb sudo dpkg -i cuda-repo-ubuntu1804-10-2-local-10.2.89-440.33.01_1.0-1_amd64.deb sudo apt-key add /var/cuda-repo-10-2-local-10.2.89-440.33.01/7fa2af80.pub sudo apt-get update sudo apt-get -y install cuda echo 'export PATH=/usr/local/cuda/bin:$PATH' >> ~/.bashrc echo 'export LD_LIBRARY_PATH=/usr/local/cuda/lib64:LD_LIBRARY_PATH' >> ~/.bashrc source ~/.bashrc
2. Install cuDNN (we recommend version v7.6.5 (November 18th, 2019), for CUDA 10.2) Go to NVIDIA's site and download .deb package. After downloading .deb package install it: bash sudo dpkg -i libcudnn7_7.6.5.32-1+cuda10.2_amd64.deb sudo dpkg -i libcudnn7-dev_7.6.5.32-1+cuda10.2_amd64.deb sudo dpkg -i libcudnn7-doc_7.6.5.32-1+cuda10.2_amd64.deb Do not forget to check if cuDNN installed properly: bash cp -r /usr/src/cudnn_samples_v7/ $HOME cd $HOME/cudnn_samples_v7/mnistCUDNN make clean && make ./mnistCUDNN cd -

3. Install AlexeyAb's fork of Darknet bash git clone https://github.com/AlexeyAB/darknet cd ./darknet # Checkout to last battle-tested commit git checkout f056fc3b6a11528fa0522a468eca1e909b7004b7 # Enable GPU acceleration sed 's/GPU=0/GPU=1/' ./Makefile # Enable cuDNN sed 's/CUDNN=0/CUDNN=1/' ./Makefile # Prepare *.so sed 's/LIBSO=0/LIBSO=1/' ./Makefile make # Copy *.so to /usr/lib + /usr/include (or /usr/local/lib + /usr/local/include) sudo cp libdarknet.so /usr/lib/libdarknet.so && sudo cp include/darknet.h /usr/include/darknet.h # sudo cp libdarknet.so /usr/local/lib/libdarknet.so && sudo cp include/darknet.h /usr/local/include/darknet.h Alternatively you can use Makefile from go-darknet repository: https://github.com/LdDl/go-darknet/blob/master/Makefile

4. Go bindings for Darknet - instructions link

5. GoCV - instructions link.
6. Blob tracking library - instructions link
7. If you want to use gRPC client-server model: gRPC - instructions link

You need to implement your gRPC server as following proto-file: https://github.com/LdDl/odam/blob/master/yolo_grpc.proto.

If you need to rebuild *.pb.go file, call this is from project root folder: protoc -I . yolo_grpc.proto --go_out=plugins=grpc:. In case of my needs I need to detect license plates on vehicles and do OCR on server-side: you can take a look on https://github.com/LdDl/license_plate_recognition for gRPC server example

After steps above done:

go install github.com/LdDl/odam/cmd/odam

Check if executable available

odam -h

and you will see something like this:

Usage of ./odam:
-settings string
        Path to application's settings (default "conf.json")

Usage

notice: targeted for Linux users (no Windows/OSX instructions currenlty)

• Prepare neural network stuff
  • Download YOLO's weights, configuration file and *.names file. Your way may warry, but here is our script: download_data.sh ./download_data_v4.sh
  • Make sure there is link to *.names file in YOLO's configuration file: [yolo] mask = 0,1,2 anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 classes=80 num=9 jitter=.3 ignore_thresh = .7 truth_thresh = 1 random=1 names = coco.names # <<========= here is the link to 'coco.names' file
• Prepare configuration file for application. Example of file: conf.json. Description of fields: ```Makefile { "video_settings": { # Video input settings "source": "rtsp://127.0.0.1:554/h264", # Link to RTSP stream "width": 1920, # Width of image in video source "height": 1080, # Height of image in video source "reduced_width": 640, # Desired width of image (for imshow and MJPEG streaming, also reduces inference time (processing > accuracy) for neural network) "reduced_height": 360, # Desired height of image (for imshow and MJPEG streaming, also reduces inference time (processing > accuracy) for neural network) "camera_id": "f2abe45e-aad8-40a2-a3b7-0c610c0f3dda" # Unique ID for video source (useful for 'client-server' model) }, "neural_network_settings": { # YOLO neural network settings "darknet_cfg": "yolov3.cfg", # Path to configuration.file "darknet_weights": "yolov3.weights", # Path to weights wile "darknet_classes": "coco.names", # Path to *.names file (labels of objects) "conf_threshold": 0.2, # Confidence threshold "nms_threshold": 0.4, # NMS threshold (postprocessing) "target_classes": ["car", "motorbike", "bus", "train", "truck"] # What classes you want to detect (if you want to use public dataset, but ignore some classes) }, "cuda_settings":{ # CUDA settings, currently useless "enable": true # CUDA settings, currently useless }, "mjpeg_settings":{ # MJPEG streaming settings "imshow_enable": false, # Do you want to enable imshow() feature (useful for testing purposes) "enable": true, # Do you want to enable this feature? "port": 35678 # Listening port fo connections }, "grpc_settings": { # gRPC 'client-server' model settings "enable": true, # Do you want to enable this feature? "server_ip": "localhost", # gRPC server's IP "server_port": 50051 # gRPC server's listening port }, "classes_settings": [ # classes settings (according to 'target_classes' in 'neural_network_settings') { "class_name": "car", # Corresponding class label "drawing_settings": { "bbox_settings": { # Setting for bounding boxes (detected objects) "rgba": [255, 255, 0, 0], # Color of bounding box border "thickness": 2 # Thickness as is }, "centroid_settings": { # Setting for centroid of bounding boxes "rgba": [255, 0, 0, 0], # Color of circle "radius": 4, # Radius of circle "thickness": 2 # Thickness as is }, "text_settings": { # Setting for text above bounding boxes "rgba": [0, 255, 0, 0], # Text color "scale": 0.5, # Size of text "thickness": 1, # Thickness as is "font": "hershey_simplex" # Text font }, "display_object_id": true # If you want to display object identifier } }, { "class_name": "motorbike", # see "car" ref. "drawing_settings": {} # if propetry is empty, then default values are used }, { "class_name": "bus", # see "car" ref. "drawing_settings": {} # if propetry is empty, then default values are used }, { "class_name": "train", # see "car" ref. "drawing_settings": {} # if propetry is empty, then default values are used }, { "class_name": "truck", # see "car" ref. "drawing_settings": {} # if propetry is empty, then default values are used } ], "tracker_settings": { # Tracked settings "tracker_type": "simple/kalman" # Use one of supported trackers. Simple tracker should fit realy simple scenes, while Kalman should be used with complicated scenes. "max_points_in_track": 150, # Restriction for maximum points in single track (>=1). Default value 10 (in case of value less than 1) "lines_settings":[ { "line_id": 1, # Unique ID for line id (useful for 'client-server' model) "begin": [150, 800], # [X1,Y1], start point of line (usually, left side) "end": [1600, 800], # [X2,Y2], end point of line (usually, right side) "direction": "to_detector", # Direction of line (possible values: 'to_detector' and 'from_detector') "detect_classes": ["car", "motorbike", "bus", "train", "truck"], # What classes must be cropped (as detected objects) that were captured by detection line. "rgba": [