Multi-object tracking by multi-feature fusion to associate all detected boxes

Abstract

Multi-object tracking (MOT) aims to estimate object trajectory in videos using either a public or private detection approach. Current trackers trained in private mode on diverse datasets make the comparison with cutting-edge methods unfair. Instead of relying exclusively on a single feature, integrating several features is more effective as it helps in data association in various conditions such as occlusion and fast Motion. Driven by this, we propose an improvised data association method by employing public detection tracker built on a CenterNet detector. In contrast to prior arts that merely used high score boxes, we improvise the data association in two stages by considering all the detections. The first stage is enforced by fusing feature embedding, IoU, and Motion features extracted from high-score boxes, followed by unmatched trajectories associated with low-score boxes using IoU similarity in the second stage. Our approach greatly surpasses cutting-edge methods in terms of excellent track quality, fewer ID switches, and high accuracy on MOT challenge datasets. The tracklet interpolation is used as a post-processing approach to fill the gap left by missing detections that improve performance even more.