Multi-Target Tracking with Historical Appearance Matching and Visually Discriminative Training Sample Selection

Abstract

The purpose of multi-target tracking is to track multiple targets without ID-switching. Multi-target tracking is one of key algorithms in many real-world systems like CCTV or autonomous driving. Designing accurate multi-target tracking algorithm is important because failure may lead to huge accident. In current paradigm of multi-target tracking, tracking by detection, we need to connect detections frame-by-frame and create tracklets. Thus, the method to find detections of the same ID is the key part of multi-target tracking algorithm. Although, multi-target tracking has been studied for a long time, there still remains a difficulty overcoming temporal errors. Temporal error occurs when targets are occluded or noisy detections appear near the targets. In those situations, tracking may fail and various errors like drift or ID-switching occur. It is hard to overcome temporal errors only by using motion and shape information. So, we propose a historical appearance matching method to get reliable appearance similarity. From support of reliable historical appearances, it can prevent tracking failures although targets are temporally occluded or last matching information is unreliable. Joint-input siamese network is used to compare appearances. For network training, we used pedestrian samples extracted from tracking sequences. Because of falsely aligned ground-truth in tracking dataset, we filtered out useless samples from tracking dataset. Additionally, we compared it with popular re-id dataset in aspect of tracking performance. Tracking performance, especially identity consistency is highly improved by attaching our methods.