In W.I.P. stage
@w.i.p.
This project is aimed to register road traffic accidents, foreign object and other anomaly situations on roads. It is pretty simple and uses just YOLO (both traditional and Ultralytics version could be used). Some advanced techniques like 3D-CNN, LSTM's and others could be used, but I do not have that much time so PR's are very welcome.
I do have also utility for monitoring road traffic flow parameters here
Diagram could speak more than a thousand words. I'm not digging into the Background Subtraction / Object Detection or MOT (Multi-object tracking) topic since it is not a main target of this project.
Diagram is prepared via https://app.diagrams.net/. Source is here.
imshow() output:
Subscribe channel on the server:
Wait until event has come:
Important notice: it has been tested on Ubuntu 22.04.3 LTS only!
It is needed to compile it via Rust programming language compiler (here is docs: https://www.rust-lang.org/learn/get-started) currently. If further I'll make a Dockerfile for CPU atleast and may be for GPU/CUDA.
Compilation from source code:
git clone https://github.com/LdDl/road-anomaly-detection
cd road-anomaly-detection
cargo build --releasePrepare neural network for detecting anomaly events.
Prepare configuration file. Example could be found here - data/conf.toml. In my example I use YOLOv8 trained on just two classes: "moderate_accident", "severe_accident".
Run:
export ROAD_ANOMALY_CONFIG=$(pwd)/data/conf.toml
./target/release/road-anomaly-detector $ROAD_ANOMALY_CONFIG
When events published into to the reciever server than you can expect following JSON structure:
{
"id": "Event identifier represented as UUID v4",
"event_registered_at": UTC UnixTimestamp when event has been registered,
"event_image": "base64 representation of an image",
"object_id": "Detection identifier. Most of time would be represented as UUID v4",
"object_registered_at": UTC UnixTimestamp when detection has been registered,
"object_lifetime": Number of second while the detection was considered "relevant",
"object_bbox": {
"x": X-coordinate of the left top of the detection,
"y": Y-coordinate of the left top of the detection,
"width": Width of the corresponding detection bounding box,
"height": Height of the corresponding detection bounding box
},
"object_poi": {
"x": X-coordinate of the center of the detection bounding box,
"y": Y-coordinate of the center of the detection bounding box
},
"object_classname": "Label for the class",
"object_confidence": Confidence that detection is classified as corresponding class label,
"zone_id": "Unique identifier for zone of interests",
"equipment_id": "Optional application name (could be considered as equipment identifier for embedded devices)"
}- Make REST API to extract and to mutate configuration;
- Make MJPEG export;
- Make publishing to custom REST API via POST request;
- Prepare some pre-trained neural networks;
- MOG2 - https://docs.opencv.org/4.x/d1/dc5/tutorial_background_subtraction.html
- MOT (Multi-object tracking) in Rust programming language - https://github.com/LdDl/mot-rs
- OpenCV's bindings - https://github.com/twistedfall/opencv-rust
- Object detection in Rust programming language via YOLO - https://github.com/LdDl/object-detection-opencv-rust
- YOLO v3 paper - https://arxiv.org/abs/1804.02767, Joseph Redmon, Ali Farhadi
- YOLO v4 paper - https://arxiv.org/abs/2004.10934, Alexey Bochkovskiy, Chien-Yao Wang, Hong-Yuan Mark Liao
- YOLO v7 paper - https://arxiv.org/abs/2207.02696, Chien-Yao Wang, Alexey Bochkovskiy, Hong-Yuan Mark Liao
- Original Darknet YOLO repository - https://github.com/pjreddie/darknet
- Most popular fork of Darknet YOLO - https://github.com/AlexeyAB/darknet
- Developers of YOLOv8 - https://github.com/ultralytics/ultralytics. If you are aware of some original papers for YOLOv8 architecture, please contact me to mention it in this README.
Video example has been taken from here
Please cite this repository if you are using it
If you have troubles or questions please open an issue.