Forward-Seed for Depth-Based Occluder Boundary Detection in UAV Target Tracking

Occlusion disrupts visual tracking of unmanned aerial vehicle (UAV) targets and makes their reappearance difficult to predict. We propose Forward-Seed to extract the foreground occluder boundary associated with a tracked target. The method combines target-relative depth ordering, a soft forward prior, multi-ray contact seeding, front-layer geodesic growth, and internal seam refinement. The evaluation includes controlled UE4 scenes, 80 flight trials across eight real UAV scenes, and a 760-event RGB-D object-tracking subset. On the additional RGB-D test subset, the mean IoU and Boundary F1 are 0.941 and 0.869, respectively. Task-matched comparisons, component ablations, and input perturbations assess the integrated design. Combining GRU trajectory predictions with Forward-Seed boundaries produces valid exit estimates for 728 of 760 events, with a mean error of 25.8 pixels across valid estimates. In complete-pipeline comparisons with LETrack, HIPTrack, and SMSTracker, GridTrack + Forward-Seed achieves tracking-success AUCs of 0.670, 0.622, and 0.597 on UE4, real-UAV, and RGBD1K tests, respectively. It achieves the highest AUC on UE4 and real-UAV tests and ranks second to SMSTracker on RGBD1K. Its recovery rates within 0.8 s of reappearance are 91.1%, 85.4%, and 84.9%, the highest among the tested pipelines. Recovery time is measured using video timestamps and excludes processing latency. The full boundary detector requires 0.806 s per 1280 × 800 image. The results support target-conditioned boundary extraction and improved recovery in the evaluated pipelines, although broader airborne validation and faster computation are still needed.