yolov5_deepsort_tensorrt

Update!

Introduction

This repo uses YOLOv5 and DeepSORT to implement object tracking algorithm. Also using TensorRTX to convert model to engine, and deploying all code on the NVIDIA Xavier with TensorRT further.

NVIDIA Jetson Xavier NX and the X86 architecture works all be ok.

Environments

1. the X86 architecture:
   • Ubuntu20.04 or 18.04 with CUDA 10.0 and cuDNN 7.6.5
   • TensorRT 7.0.0.1
   • PyTorch 1.7.1_cu11.0, TorchVision 0.8.2+cu110, TorchAudio 0.7.2
   • OpenCV-Python 4.2
   • pycuda 2021.1
2. the NVIDIA embedded system：
   • Ubuntu18.04 with CUDA 10.2 and cuDNN 8.0.0
   • TensorRT 7.1.3.0
   • PyTorch 1.8.0 and TorchVision 0.9.0
   • OpenCV-Python 4.1.1
   • pycuda 2020.1

Speed

The following data are tested in the case of single target in the picture. the X86 architecture with GTX 2080Ti :

NVIDIA Jetson Xavier NX:

Inference

1. Clone this repo

   git clone https://github.com/cong/yolov5_deepsort_tensorrt.git
   

2. Install the requirements

   pip install -r requirements.txt
   

3. Run

   python demo_trt.py
   

   

Convert

Convert PyTorch yolov5 weights to TensorRT engine.

Notice: this repo uses YOLOv5 version 4.0 , so TensorRTX should uses version yolov5-v4.0 !

1. generate ***.wts from PyTorch with ***.pt.

   git clone -b v4.0 https://github.com/ultralytics/yolov5.git
   git clone -b v4.0 https://github.com/wang-xinyu/tensorrtx.git
   # download https://github.com/ultralytics/yolov5/releases/download/v5.0/yolov5s.pt
   cp {tensorrtx}/yolov5/gen_wts.py {ultralytics}/yolov5
   cd {ultralytics}/yolov5
   python gen_wts.py yolov5s.pt
   # a file 'yolov5s.wts' will be generated.
   

2. build t{tensorrtx}/yolov5 and generate ***.engine

   cd {tensorrtx}/yolov5/
   # update CLASS_NUM in yololayer.h if your model is trained on custom dataset
   mkdir build
   cd build
   cp {ultralytics}/yolov5/yolov5s.wts {tensorrtx}/yolov5/build
   cmake ..
   make
   # serialize model to plan file
   sudo ./yolov5 -s [.wts] [.engine] [s/m/l/x/s6/m6/l6/x6 or c/c6 gd gw]
   # deserialize and run inference, the images in [image folder] will be processed.
   sudo ./yolov5 -d [.engine] [image folder]
   # For example yolov5s
   sudo ./yolov5 -s yolov5s.wts yolov5s.engine s
   sudo ./yolov5 -d yolov5s.engine ../samples
   # For example Custom model with depth_multiple=0.17, width_multiple=0.25 in yolov5.yaml
   sudo ./yolov5 -s yolov5_custom.wts yolov5.engine c 0.17 0.25
   sudo ./yolov5 -d yolov5.engine ../samples
   

3. Once the images generated, as follows. _zidane.jpg and _bus.jpg, convert completed!

Convert PyTorch DeepSORT weights to TensorRT engine.

1. generate ***.onnx from PyTorch with ***.pt.

   git clone https://github.com/ZQPei/deep_sort_pytorch
   git clone https://github.com/GesilaA/deepsort_tensorrt.git
   # 
   cp {GesilaA}/deepsort_tensorrt/exportOnnx.py {ZQPei}/deep_sort_pytorch
   cd {ZQPei}/deep_sort_pytorch
   python exportOnnx.py
   # a file 'deepsort.onnx' will be generated.
   cp {ZQPei}/deep_sort_pytorch/deepsort.onnx {GesilaA}/deepsort_tensorrt
   

2. build {GesilaA}/deepsort_tensorrt and generate ***.engine

   cd {GesilaA}/deepsort_tensorrt
   # 
   mkdir build
   cd build
   cmake ..
   make
   # serialize model to plan file
   ./onnx2engine ../resources/deepsort.onnx ../resources/deepsort.engine
   # test
   ./demo ../resources/deepsort.engine ../resources/track.txt
   

Customize

1. Training your own model.
2. Convert your own model to engine(TensorRTX's version must same as YOLOV5's version).
3. Replace the ***.engine and libmyplugins.so file.

Optional setting

• Your likes are my motivation to update the project, if you feel that it is helpful to you, please give me a star. Thx! :)
• For more information you can visit the Blog.