chenhanpolyu / DynTrack

yolo_fdsst_piv

A universal object detecting, tracking, and motion estimating framework

Install dependencies

source install_deps.sh

Demo:

RGB image based object tracker:

Object tracking and motion estimation in 3D space:

If we use PIV method only to track and classify the objects into dynamic or static by velocity, the classification is often wrong due to the error in velocity estimation:

launch

Build this repo in your /src folder, source the workspace.

launch the object detector and tracker on RGB image: roslaunch yolodetector yolo_ros.launch Please modify the related parameters in the config/yolo_ros.yaml. Set if_debug=true if you want to check the tracking bounding box in the image output.

The object velocity estimator based on point cloud is also launched in the same launch file:

It is developed from 2d PIV method, I make it applicable for 3d point cloud and design two simple local features to make the velocity estimation more accurate. It can classify the dynamic and static objects with the help of the image-based object detector, and only output the velocity of the dynamic ones.