The c++ version of https://github.com/nwojke/deep_sort.

• opencv3
• tensorflow 1.4
• cuda 8.0
• cudnn 6.0

develop tool: Nsight Eclipse Edition Version:9.1

Defore you run, you need:

1. download YOLOV2's related model files https://pjreddie.com. Copy the cfg, data directory and yolo.weights to the RUNNINGDATA/darknet_data.
2. download the CNN checkpoint file resources/networks/mars-small128.ckpt-68577 and put to RUNNINGDATA/tensor_networks.

cd Debug
make
cd .. #into the project root directory
./Debug/VideoTracking

There are there options in main.cpp which you cann't define simutanously. They are RUNGT RUNMOTTENSOR and RUNLOCALVIDEO.

• RUNGT: you need download MOT dataset and define the specific directory by MOTDIR in the main.cpp. FUNCTION: tracking mot dataset using mot groundtruth.
• RUNMOTTENSOR: like RUNGT. FUNCTION: tracking mot dataset using mot detection groundtruth but generate feature by tensorflow.
• RUNLOCALVIDEO: you need define the local video's path by VIDEO in the main.cpp. FUNCTION: tracking local video without any known data.