Lightweight Small-Object Detection for Urban UAV Imagery with Content-Aware Feature Reconstruction and Gradient-Adaptive Localization

Abstract

Small-object detection in unmanned aerial vehicle (UAV) imagery is hindered by limited target pixels, dense spatial distributions, background interference, and scale-sensitive bounding-box regression. This study develops a lightweight YOLO11n configuration in which established SCSA recalibration, DySample reconstruction, decoupled prediction, and SimOTA assignment act at successive stages of the detection pipeline. Its principal methodological contribution is a gradient-adaptive WIoU–NWD objective that uses previously observed regression-gradient fluctuations to balance overlap-oriented and distribution-based localization without altering the inference graph. On the official VisDrone2019-DET test-dev server, the 2.48 M-parameter model achieves 45.9% mAP@0.5, 27.0% mAP@0.5:0.95, and 20.2% APsmall, improving the YOLO11n baseline by 2.3, 1.4, and 2.4 percentage points, respectively; it also improves mAP@0.5/mAP@0.5:0.95 by 2.2/1.3 points on the vehicle-focused UAVDT benchmark and reaches 40.2 FPS on a Jetson Orin NX in 15 W mode using TensorRT FP16. The two benchmarks mainly represent urban, traffic, and low-altitude surveillance imagery; consequently, the cross-dataset result supports transfer within these conditions rather than universal generalization to all UAV applications. These results support a compact single-pass accuracy–efficiency trade-off for resource-constrained UAV perception, while the modest margin over fixed loss weighting indicates that the adaptive mechanism should be interpreted as an incremental, primarily small-object localization improvement rather than a complete solution to regression instability.