Real time multiple object tracking with improved object interference handling

Abstract

Having the ability to track multiple objects in video sequences by object detection can result in not only useful applications such as video surveillance but also problems which includes object re-identification, poor motion prediction and occlusion. While multi-object tracking (MOT) is a deeply explored area, it is not yet successfully solved using computer vision methods and artificial intelligence. Due to highly dynamic environment in MOT, objects often collide and occlude each other hence resulting in false initializing and ID switching. In order to reduce the occurrence of these issues, this paper presents a simple method which integrates existing tracker and various methods such as calculating the cosine differences in visual features and changing the relevant thresholds, to improve the accuracy. This tracking system is evaluated on MOT17 benchmark detection rubrics and the results achieved showed improvements as compared to having no such extensions.