Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/emptysoal/TensorRT-YOLOv8-ByteTrack
An object tracking project with YOLOv8 and ByteTrack, speed up by C++ and TensorRT.
https://github.com/emptysoal/TensorRT-YOLOv8-ByteTrack
bytetracker object-detector object-tracking tensorrt yolov8
Last synced: 4 months ago
JSON representation
An object tracking project with YOLOv8 and ByteTrack, speed up by C++ and TensorRT.
- Host: GitHub
- URL: https://github.com/emptysoal/TensorRT-YOLOv8-ByteTrack
- Owner: emptysoal
- License: mit
- Created: 2023-10-24T09:28:36.000Z (about 1 year ago)
- Default Branch: main
- Last Pushed: 2024-07-12T03:43:45.000Z (4 months ago)
- Last Synced: 2024-07-12T04:34:44.129Z (4 months ago)
- Topics: bytetracker, object-detector, object-tracking, tensorrt, yolov8
- Language: C++
- Homepage:
- Size: 45.3 MB
- Stars: 87
- Watchers: 1
- Forks: 10
- Open Issues: 0
-
Metadata Files:
- Readme: README-en.md
- License: LICENSE
Awesome Lists containing this project
- awesome-yolo-object-detection - emptysoal/TensorRT-YOLOv8-ByteTrack - YOLOv8-ByteTrack?style=social"/> : An object tracking project with YOLOv8 and ByteTrack, speed up by C++ and TensorRT. (Lighter and Deployment Frameworks)
- awesome-yolo-object-detection - emptysoal/TensorRT-YOLOv8-ByteTrack - YOLOv8-ByteTrack?style=social"/> : An object tracking project with YOLOv8 and ByteTrack, speed up by C++ and TensorRT. (Lighter and Deployment Frameworks)
README
# TensorRT C++ api deploy YOLOv8 + ByteTrack
## Introduction
- Based on `TensorRT-v8` , deploy `YOLOv8` + `ByteTrack` ;
- Support `Jetson` series, also `Linux x86_64`;
Main work I have done:
1. Refer to [tensorrtx](https://github.com/wang-xinyu/tensorrtx/tree/master/yolov8) ，model: `.pth` -> `.engine`,extract the inference part of the code, encapsulated into C++ classes, easy to call other projects ;
2. Preprocessing replaced with my own CUDA programming preprocessing;
3. Post-processing removed CUDA programming because it was not significantly faster in tests compared to CPU post-processing ;
4. The post-processed NMS greatly reduces conf_thres hyperparameters due to the principle of `ByteTrack` tracking, which is very important ;
5. `YOLOv8` inference compiles to a dynamic link library to decouple projects;
6. Reference official [ByteTrack TensorRT deploy](https://github.com/ifzhang/ByteTrack/tree/main/deploy/TensorRT/cpp) , modify its interface to the `YOLO` detector;
7. `ByteTrack` also compiles to a dynamic link library, further decoupling projects;
8. Add category filtering function, you can set the category you want to track in `main.cpp` line 8 .
## Effect
# Environment
1. Base requirements：
- `TensorRT 8.0+`
- `OpenCV 3.4.0+`
2. My running environment on `Jetson Nano` is as follows:
- The burned system image is `Jetpack 4.6.1`，original environment is as follows：
| CUDA | cuDNN | TensorRT | OpenCV |
| ---- | ----- | -------- | ------ |
| 10.2 | 8.2 | 8.2.1 | 4.1.1 |
- Install Eigen
```bash
apt install libeigen3-dev
```
## Model conversion
Get the serialized file of TensorRT, suffix.engine
- First get the wts format model file, link: [yolov8s.wts](https://pan.baidu.com/s/16d_MqVlUxnjOhLxVyjQy8w) , code：gsqm
- Then follow these steps:
```bash
cd {TensorRT-YOLOv8-ByteTrack}/tensorrtx-yolov8/
mkdir build
cd build
cp {path/to/yolov8s.wts} .
cmake ..
make
./yolov8 -s yolov8s.wts yolov8s.engine s
cd ../../
mkdir yolo/engine
cp tensorrtx-yolov8/build/yolov8s.engine yolo/engine
```
## Run tracking
- Follow these steps
```bash
mkdir build
cd build
cmake ..
make
./main ../videos/demo.mp4 # The path to your own video
```
Then the `result.mp4` will be in the build directory, is to track the effect of the video file
If you want the tracked video to play in real time, you can uncomment line 94 of main.cpp.
# Reference
- [tensorrtx](https://github.com/wang-xinyu/tensorrtx/tree/master/yolov8)
- [ByteTrack](https://github.com/ifzhang/ByteTrack)