PMR-TransT:Pedestrian Target Tracking Based on Mutual Refinement

Pedestrian tracking faces critical challenges in real-world scenarios, primarily due to drastic pose variations, frequent target occlusion, and complex background interference. Conventional methods often lack the representational capacity and robustness needed for practical deployment. To address these issues, we propose PMR-TransT, a novel Transformer-based pedestrian tracking framework. It features a Mutual Refinement (MR) module that enhances semantic consistency between template and search branches through fine-grained, cross-attention-based information interaction, effectively suppressing background noise and highlighting target-specific features. Additionally, we introduce a DBSCAN-based pseudo-label generation mechanism combined with a triplet loss constraint to automatically mine feature-consistent samples from unlabeled data and optimize feature learning. Experimental results demonstrate the effectiveness of our approach. PMR-TransT achieves an Average Overlap (AO) improvement of 0.3 on the GOT-10k dataset and delivers superior performance on the LaSOT dataset compared to state-of-the-art baseline methods, validating its robust practical application value. Ultimately, these improvements provide a highly reliable solution, validating its substantial application value for intelligent surveillance, autonomous driving, and humancomputer interaction systems.