Accuracy and Long-Term Tracking via Overlap Maximization Integrated with Motion Continuity

Abstract

The baseline is ATOM which aims at solving the problem of accurate target state estimation by proposing a novel tracking architecture. The architecture consists of dedicated target estimation and classification components. Classification component is trained online to guarantee high discriminative power in the presence of distractors. Target estimation is performed by the IoU-predictor network inspired by the IoU-Net which was recently proposed for object detection as an alternative to typical anchor-based bounding box regression techniques. In this work, we further enhance the performance of ATOM by embedding Squeeze-and-Excitation (SE) blocks into IoU-Net in ATOM to recalibrate useful features and suppress useless features and obtain ATOMFR. To solve the abnormal changes in the target box in ATOMFR, we add the Relocation Module on ATOMFR and get ATOMFR (RL). To solve the occlusion problem, we introduce the Inference Module into ATOMFR (RL) and obtain ATOMFR (RL + InF). Experimental results on VisDrone2019-SOT test set demonstrate the state-of-the-art performance of ATOMFR (RL + InF) compared with several existed trackers and it ranks the second place among all competitors.