Robust Hierarchical Data Association for Multiple Object Tracking under Abrupt Appearance Changes
Sprache des Titels:
Proc. of the International Conference on Electrical, Computer and Energy Technologies (ICECET)
In automated traffic surveillance, the correct interpretation of a perceived scene from a stream of video data combined with CNN?based object detection is an emerging topic. To understand and predict the scene, detected vehicles must be re?identified in a stream of consecutive video frames, and their trajectories need to be reconstructed. Major challenges in video? based Multiple Object Tracking (MOT) are abrupt appearance changes of an object, which e.g. are caused by perspective changes and mutual occlusion of objects. This contribution presents a hierarchical data association approach with error? compensating association rounds, which is combined with a model?based trajectory tracking approach. The benefit of the proposed approach is evaluated using HOTA-metrics (Higher Order Tracking Accuracy).