TdmTracker: Multi-Object Tracker Guided by Trajectory Distribution Map

Abstract

With the great progress of object detection, some detection-based multiple object tracking (MOT) paradigms begin to emerge, including tracking-by-detection, joint detection and tracking, and attention mechanism-based MOT. Due to the separately executed detection, embedding, and data association, tracking-by-detection-based methods are much less efficient than other end-to-end MOT methods. Therefore, recent works are devoted to integrating these separate processes into an end-to-end paradigm. Some of the transformer-based end-to-end methods introducing track queries to detect targets have achieved good results. Self-attention and track query of these methods has given us some inspiration. Moreover, we adopt optimized class query instead of static learned object query to detect new-coming objects of target category. In this work, we present a novel anchor-free attention mechanism-based end-to-end model TdmTracker, where we propose a trajectory distribution map to guide position prediction, and introduce an adaptive query embedding set and query-key attention mechanism to detect tracked objects in the current frame. The experimental results on MOT17 dataset show that the TdmTracker achieves a good speed-accuracy trade-off compared with other state-of-the-arts.