Learning Distance Constrained Transformation for Video Tracking in Car-Following

Recent advances in video tracking with discriminative correlation filters leverage diverse observation models. However, fusing hand-crafted and deep convolutional neural network representations equivalently would overly constrain resolution conditions for template matching, leading to peak response slippage and jittery neighboring search processes, especially problematic in autonomous driving scenarios. This article addresses the inference conservatism issue in multitype feature tracking. We propose a target-observation constraint framework to formalize discrimination conservatism across feature map channels. A learning constraint transformation methodology is introduced to cluster similar representations while pushing dissimilar ones apart. These discriminant constraints are further fine-tuned through joint learning with correlation filters, improving the positional precision of detection responses. Additionally, we propose an updating strategy that suppresses low scores of symmetric dispersion ratio, enhancing tracking robustness. Extensive evaluations on five tracking datasets demonstrate the superior performance of our approach: UAV20L, UAVDT, OTB-100, VOT-2019, and LaSOT.