Robust pedestrian tracking in video sequences using an improved STF module

Object tracking technology based on image processing has made great progress recently. Based on the track-by-detection framework, the tracking algorithms are often combined with deep neural networks to perform online target tracking. However, existing motion models assume linearity and are sensitive to sudden changes in trajectories due to occlusion, overlap, or other detection issues. In this paper, we modified the existing object tracking algorithms and introduced a strong tracking filter (STF) module to the track-by-detection framework for solving the sudden change problem of the target. The STF module is designed to have a strong ability to track sudden changes by orthogonalizing the residual sequence. When the trajectory of the target is stable, the STF module returns to the inactive state, behaving similarly to tracking algorithms that follow conventional linear models. Experimental results on a public pedestrian tracking dataset show that the proposed method improves tracking performance on various metrics, including the ability to rematch missed trajectories. Moreover, compared with existing algorithms, the proposed method exhibits strong stability under noisy conditions.