Graph Convolutional Network-based Centre-Object Tracking for the multi-dynamic agents for Autonomous Vehicles

Abstract

Multi-Object tracking goals to localize, classify and track all object instances of each class throughout an image sequence. It is very useful to understand the video scenes and very desirable for computer vision-based applications that require detailed pixel-level information, such as video editing and autonomous driving. Our method involves Detection-based tracking, which generally requires detected objects with a tracker that links the boundary boxes (bboxes) to generate target trajectory. To optimize the similarity between tracked and detected objects are the basic part of multi-object tracking (MOT) frameworks. To mark this issue, ongoing works concurrently enhance the matching technique of data modules with joint MOT structure, which has expressed improved execution. The key thought behind our GCNN approach is to take into consideration and learn the distinctive or selective features for local and global matching. It takes into account the relations between nearby objects in a frame sequence and thus provides a better tracking performance as compared to the existing models. We demonstrated our model on the MOT datasets that showed the viability of our GCNN-based joint MOT approach which includes the Hungarian algorithm to assign the identity of the object in the next frame. © 2021 IEEE.