Extending Simple Online and Realtime Tracking with Sparse Optical Flow

Abstract

Detection and tracking of objects in video data is a common topic in computer vision and widely used for traffic surveillance. Simple Online and Realtime Tracking (SORT) is a Multi-Object Tracking (MOT) algorithm which performs in realtime by using efficient algorithms. SORT’s tracking performance is based on detection quality and may suffer due to detection failures. This work proposes an extension to SORT’s Kalman filter measurement model with Robust Local Optical Flow to add a velocity component to the filter’s update vector and improve robustness against detection failures and decrease track id switches. The elaborated extension, termed SORT_OF, is tested and benchmarked on parts of the MOT15 tracking benchmark as well as on a self created benchmark representing a DLR real world scenario. The proposed tracking method still works in realtime.