Enhanced Multi-Scale Feature Extraction for Small Vehicle Detection from UAV Perspective

Abstract

Small vehicle detection in uncrewed aerial vehicle (UAV) imagery is severely hindered by low target resolution, drastic scale variations, and complex background occlusion, yielding high false-negative rates in conventional detectors. Here, we propose an enhanced YOLO11 architecture optimized to resolve these limitations. To prevent the dilution of small-object features without inflating parameter counts, we replace the standard spatial pyramid pooling module with a streamlined variant (SimSPPF). We further integrate a coordinate attention mechanism to explicitly embed positional data into channel dependencies, enhancing spatial sensitivity. Finally, an adaptively spatial feature fusion (ASFF) head is deployed to resolve gradient conflicts between deep semantic and shallow positional features. Evaluated on the VisDrone benchmark, our model achieves a mean average precision (mAP50) of 35.4%, outperforming the baseline by 2.4 percentage points. This synergistic architecture robustly mitigates missed detections and false positives in cluttered traffic environments.