Reverse Procedure for Lost Track Recovery in Multi-Object Tracking

Abstract

Multi-object tracking (MOT) performance depends strongly on object-detection quality, particularly in tracking-by-detection systems. Track loss occurs when the detector misses a target in one or more video frames. The detector's performance may also affect the data association process, and the number of false negatives may increase drastically. Many studies have focused on improving detector performance, while other work has aimed to improve data association strategies. However, track loss remains a persistent problem when a target is missed for several consecutive frames. To address the problem, we propose a lost track recovery method based on reverse processing. We propose a method that combines forward and backward processing to recover missing trajectories and decrease false negatives. In forward processing, a Kalman filter is used for predicting the state of all tracklets, including, but not limited to, temporarily lost objects. Backward processing is triggered when a previously lost target is subsequently re-detected, and the recovery condition is evaluated using information from neighbouring objects. By using forward state prediction and backward trajectory recovery, the method reconstructs eligible missing trajectory states. On MOT17, the proposed method improves MOTA from 80.3% to 80.9%, IDF1 from 77.3% to 79.9%, and HOTA from 63.1% to 64.9%. On MOT20, IDF1 improves from 75.2% to 77.2% and HOTA from 61.3% to 63.2%, while MOTA changes slightly from 77.8% to 77.7%. These results show that the recovery procedure improves selected tracking metrics, while its benefit is less consistent in very dense scenes.