Abstract
Event-based UAV small-target segmentation requires recovering sparse and incomplete foreground events from structured background activity. Longer observation windows provide additional target evidence but also accumulate clutter, making temporal feature integration critical for accurate segmentation. We propose an Evolution-Aware Event Segmentation Network (EAES-Net) that models local transitions between adjacent feature states within a sparse encoder–decoder. The Temporal Evolution Encoder (TEE) jointly encodes neighboring states and their difference to construct transition representations. Local Evolution Consistency Aggregation (LECA) evaluates the cosine agreement of successive transitions at each spatial location and aggregates evolution evidence using spatially varying weights. Evolution-Guided Refinement (EGR) uses a learned gate to combine evolution-aware information with the original sparse features before decoding. A foreground consistency loss further regularizes adjacent target representations during training. Experiments on the EV-UAV benchmark demonstrate improvements in foreground-event segmentation, target detection, and false-alarm suppression over the evaluated reference methods. Component and design ablations quantify the contributions of the proposed modules, while temporal sensitivity experiments examine performance across different observation windows and temporal resolutions.