DCGOFTracker: Dual-Cue Guidance and Occlusion-Friendly for Multiobject Tracking in Remote Sensing Videos

Mainstream multi-object tracking (MOT) algorithms typically employ a joint detection and association approach. However, in remote sensing videos characterized by crowded scenes with numerous small objects and complex backgrounds, existing trackers often suffer from high rates of missed and false detections. To address these challenges, this study proposes a novel joint detection and tracking (JDT) framework: the dual cue guided and occlusion friendly tracker (DCGOFTracker). This framework integrates local detection enhancement for small targets and refines the association matching process to improve tracking performance. Specifically, it first incorporates local motion cues into the detection stage through a motion-guided feature warping (MFW) module. Then, the inter-frame enhancement (IFE) module further improves the feature representation of small targets using inter-frame correlation cues. Finally, an occlusion-aware offset-based association (OAOB) branch is introduced in the data association matching phase to enhance tracking robustness. We validated the effectiveness and robustness of DCGOFTracker on two datasets. On the UAVDT dataset, the model achieved a multiple object tracking accuracy (MOTA) of 55% and an identity F1 score (IDF1) of 62.9%. Meanwhile, on the satellite dataset, the MOTA and IDF1 scores reached 75.5% and 83%, respectively. These results demonstrate that the model exhibits strong generalization capabilities and can accurately and efficiently track small objects in remote sensing image.