Adaptive Temporal Reallocation and Trajectory-Aware Modulation for Event-Level Segmentation of Small-Scale UAVs

Event-level segmentation of small-scale unmanned aerial vehicles (UAVs) is challenging because target-generated events are sparse and are easily mixed with background motion and sensor noise. This study extends EV-SpSegNet with Adaptive Temporal Reallocation (ATR), Trajectory-Aware Modulation (TAM), and Directional Mask Regularization (DMR). ATR reallocates the temporal coordinates used for sparse voxel construction according to interval-wise spatiotemporal linearity while preserving the original events, features, and labels. TAM combines x-t and y-t directional features and selectively modulates intermediate feature responses, while DMR discourages excessive positive modulation during training. On the EV-UAV benchmark, the proposed method improves the IoU of the reproduced EV-SpSegNet baseline from 57.10% to 65.74% and reduces Fa from 2.27×10−4 to 0.61×10−4. Additional evaluation on NeRDD, multi-seed training, and controlled ablation analyses further demonstrate the effectiveness of the proposed approach.