ADAPT-Tracker: Adaptive dynamic perception and trajectory-consistent network for aerial multi-object tracking

Aerial Multi-object tracking is hindered by coupled dynamics in which rapid UAV ego-motion disrupts inter-frame alignment and stochastic target motion amplifies non-linear observation noise, jointly breaking temporal consistency—especially for tiny objects. Existing trackers struggle in this compound regime because convolutional backbones suppress small-target cues under clutter, while constant-velocity motion models are brittle to maneuver-induced jitter and intermittent detections. We propose ADAPT-Tracker, a unified framework that couples scale-aware representation learning with high-order geometric trajectory refinement. On the perception side, Motion-Aware Receptive Field Modulation (MRFM) decomposes features into complementary granularity components to preserve fine-grained target details while suppressing distractors, and performs structure-preserving channel mixing to strengthen small-object representations. On the motion side, High-Order Geometric Stabilized Trajectory Refinement (H-GSTR) conducts bi-directional trajectory inpainting to restore continuity under occlusion and missed detections, and applies curvature-regularized kinematics to stabilize maneuver-induced jitter. Beyond overall metrics, a sequence-level recovery study on UAVDT shows that H-GSTR recovers 8.66% additional valid detections on average, peaking at 27.27% in high-viewpoint small-target scenes. Extensive experiments on UAVDT and VisDrone-MOT demonstrate robustness under scale variation and abrupt motion, improving over the strongest competing method by approximately +2.00 (HOTA), +3.00 (MOTA), and +1.50 (IDF1) percentage points on VisDrone-MOT. Consistent superiority is observed on the UAVDT benchmark.