This project implement the Multi-Object-Tracking(MOT) base on SOLOv2 and DeepSORT with C++。 The instance segmentation model SOLOv2 has deploy to TensorRT, and the postprocess implement with Libtorch. Therefore, the frame rate of detection and tracking can exceed 40 FPS。 Test video was showed here

• Ubuntu
• Cuda10.2
• cudnn8
• GCC >=9
• TensorRT8
• Opencv3.4
• Libtorch1.8.2
• CMake3.20

SOLO
SOLOv2.tensorRT
Yolov5_DeepSort_Pytorch
libtorch-yolov3-deepsort

see Solov2-TensorRT-CPP

this part is base the libtorch-yolov3-deepsort . Download the deepsort model ckpt.t7 from here. Then use the script conv_model_format.py convert model format from ckpt.t7 to ckpt.bin.

Firstly edit the config.yaml to right setting. Then compile the project:

mkdir build && cd build
cmake ..

Run the demo

./tracking ../config/config.yaml