Frequency-Guided Feature Representation for Instance Segmentation in Aerial View Traffic Accident Scenes

Accurate instance-level perception of aerial view traffic accident scenes is the foundation of accident investigation. However, existing accident datasets mainly support event-level video analysis or coarse spatial localization, and rarely provide pixel-level vehicle masks together with accident-involved labels. To address this gap, we construct the Drone-oriented Accident Recognition and Segmentation (DARS) dataset, an instance segmentation dataset specifically developed for aerial view traffic accident scenes. DARS contains 7603 images and 49,613 vehicle instances with six classes jointly defined by vehicle type and accident-involved status. Statistical analysis reveals the class distribution, scale variation, and accident-type composition of the dataset. We further introduce a Frequency-Guided Adaptive Downsampling (FGAD) method into YOLO26n-seg to improve hierarchical feature extraction. FGAD performs content-adaptive aggregation of spatial candidate features, while wavelet-derived frequency information guides candidate weight estimation and provides an additional residual pathway. On DARS, the proposed method achieves 52.37% recall, 51.24% mAP@50, 41.18% mAP@75, and 36.77% mAP@50–95, outperforming other methods in terms of instance segmentation accuracy. On a UAV-based Vehicle Segmentation Dataset (UVSD), it consistently improves over YOLO26n-seg, reaching 75.61%, 57.74%, and 44.60%, respectively. These results support DARS as an instance-level benchmark and demonstrate the applicability of FGAD to aerial traffic perception, providing a basis for fine-grained accident scene analysis in intelligent and sustainable transportation systems.