English | 简体中文
- This project is a C++ implementation of Yolo-DeepSORT using TensorRT for inference;
- Provides a Dockerfile for quick setup of the development environment;
- Simply provide the ONNX file, and when creating the model instance, it will automatically parse the ONNX and serialize the engine file (*.trtmodel) to the workspace directory;
- My other PyTorch version implementation, including line-crossing detection for pedestrian counting: Yolov5_Deepsort_Person_Count
Refer to the README to use the Docker container, or configure manually as follows:
- python: 3.8
- cuda: 11.2
- cudnn: 8.2.2.26
- tensorRT: 8.0.3.4
- protobuf: 3.11.4
Create a workspace directory:
mkdir workspaceDownload files to the workspace directory:
| File | Link |
|---|---|
| onnx, test.mp4 | Download (code: zxao) |
- yolov5s.onnx is exported from yolov5-6.0, and the ONNX model for deepsort is exported from deep_sort_pytorch. Refer to the script exportOnnx.py for more details.
- ptq_yolov5s.onnx is a quantized model. For reference, see https://github.com/Tongkaio/yolov5_quant.
Modify the relevant header and library paths in the MakeFile (if you are using a container created from the provided Dockerfile, this step is not needed), then execute:
make runInference results will be displayed during runtime, press ESC to exit.
Currently, on a GeForce RTX 2060, the inference speed for test.mp4 is approximately 40 ms/frame.
-
Infer, Yolo, and DeepSORT are encapsulated using interface mode and RAII:
- infer.h， yolo.h，deepsort.h only expose
create_*and inference interfaces. - Use
create_*to create object instances, which will automatically parse the ONNX file, generate, and load the engine.
- infer.h， yolo.h，deepsort.h only expose
-
infer.cpp: Divides into four threads, with producer-consumer relationships between each pair:
