Learning Modality Residual Regularized Correlation Filter for Real Time RGB-T Tracking

With the development of image sensor technology, RGB-T tracking methods have demonstrated their potential performances in visual object tracking tasks. In order to effectively utilize the heterogeneity of visible images and thermal infrared images and handle the challenging scenarios (e.g., thermal cross, deformation and partial occlusion, etc.), in which several existing methods are limited in, a novel tracking framework using handicraft and tiny deep features is proposed in this paper. First, by integrating the residue of features extracted from two modalities, a modality residual regularization term is generated in regression of the correlation filter (CF) framework to learn more distinct features. Second, a progressive temporal regularization is introduced as well to handle the rapid object appearance variations. Moreover, a simple but effective dynamic model update and redetection mechanism is designed to further improve the robustness of the proposed tracker in complex scenarios. Exhaustive experiments are conducted on 284 challenging sequences on two tracking benchmarks, (GTOT and RGBT234), and proves the effectiveness of the proposed tracker in accuracy, robustness. Besides, the proposed tracking method has reached 65.34 FPS, which is faster than mainstream trackers.