MHA-Track: Multi-Horizon Association for Vehicle Tracking in Low-Frame-Rate Aerial Imagery

Kurzfassung

Vehicle tracking in aerial image sequences is promising for traffic monitoring, offering wide coverage and high resolution, but low framerates and large camera motions break standard tracking assump tions. Small, similar vehicles and frequent unmatched cases make adjacent frame appearance-based association unreliable. We introduce MHA Track, a multi-horizon GNN-based tracker designed for low framerate aerial imagery. For each vehicle and anchor frame, MHA-Track scores candidate matches across six temporal horizons within overlap ping windows of eight camera-motion-aligned frames, competing matches against learned continue and terminate states to handle disappearances. A Gaussian motion model predicts per-horizon displacements, and a two pass inference refines associations using trajectory speed estimates. Aggregating evidence across horizons and windows enables robust tracking under occlusion and large displacements. On the VETRA dataset, MHA-Track outperforms the best baseline by +1.94 HOTA, +1.65 MOTA, +2.97 IDF1, and halves identity switches. MHA-Track excels at sharp turns, large camera motions, recovery after occlusion, and tracking similar vehicles in close proximity. Code is publicly available at https://github.com/ManuelM95/MHA-Track.