This repository uses yolov5 and TensorRT to deploy object tracking algorithms. Deep-sort algorithm are used to track the objects.Thank you all for the contribution of tensorrtx and deepsort.This project is all implemented in C++ language, and the preprocessing code of the detection part is run on the GPU using npp programming.The cost time of the inference is 3ms, and the preprocessing time is 6ms.
Tesla T4 cuda 11.1 TensorRT 7.2.2.3
git clone https://github.com/xuyufeng1995/yolov5_deepsort_tensorrt.git
cd tensorrt_yolov5_tracker
mkdir build
cd build
cmake ..
makeYou could change the batch size in yolov5/yolov5.h You could change the input image size in yolov5/yololayer.h
We provide the support for building models of yolov5-4.0 in our project. The default branch is corresponding to yolov5-4.0.
If other versions are wanted, please kindly refer to tensorrtx to build the engine and then copy to the folder of this project.
At first, please put the yolov5/gen_wts.py in the folder of corresponding version of Yolov5. For example, if you use the model of yolov5-4.0, please download the release of yolov5-4.0 and put the gen_wts.py in its folder. Then run the code to convert model from .pt to .wts. Then please use the following instructions to build your engines. This code will generate yolov5s engine file and filename is yolov5s.engine.
./Tracker -s # build yolov5 model
./Tracker -v ../huangxing.mp4 # run