CCDMOT: An Optimized Multi-Object Tracking Method for Unmanned Vehicles Pedestrian Tracking

Multi-object tracking (MOT) is pivotal for under-standing environments in which unmanned vehicles function. The Joint Detection and Embedding (JDE) paradigm, merging target motion and appearance for data association, stands as a cornerstone in MOT. Despite its recognition, conventional methods exhibit limitations, particularly when faced with intricate traffic settings, diverse pedestrian appearances, and recurrent occlusions. Addressing these challenges, this paper introduces CCDMOT, an optimized MOT algorithm. Inspired by FairMOT, CCDMOT incorporates a Criss-Cross Attention Block to refine feature extraction and boost the relationship between varied dimensional features. Moreover, our innovative Dynamic Re-ID Embedding network augments the Re-ID tasks, amplifying stability in congested settings and mitigating target losses. Comparative assessments on the MOT17 validation set indicate a significant 1.5 MOTA improvement over FairMOT, with ID switches reducing from 404 to 340, emphasizing CCDMOT's enhanced stability. Additionally, CCDMOT also achieved satisfactory results on the MOT challenge. Finally, qualitative evaluation on the autonomous driving dataset KITTI confirmed that the framework proposed in this study can effectively be applied to pedestrian tracking in unmanned vehicles, demonstrating the practical significance and applicability of our work.