HPSNet: A Three-Stage Enhanced YOLOv11 Detector for Person-Overboard Detection in Maritime UAV Imagery

Abstract

Person-overboard detection from maritime unmanned aerial vehicle (UAV) imagery is challenging because the targets occupy very few pixels, sea-surface clutter is severe, human appearance varies substantially, and real-time processing is required. Existing detectors therefore struggle to meet the demands of practical maritime search and rescue. This paper presents HPSNet, an accuracy- and recall-oriented detector designed for maritime UAV imagery. HPSNet uses YOLOv11 as its baseline and introduces three complementary modifications. A channel transposed attention (CTA) module is embedded in the backbone to improve the discrimination of target features from complex sea-surface interference. A Giraffe feature pyramid network (GFPN) replaces the original feature-fusion network to strengthen multiscale information exchange and preserve cues from extremely small targets. A Dynamic Head (DyHead) adapts the predictions to variations in target scale, location, and appearance. HPSNet is evaluated against 12 representative detectors on the public Person Detection in Water and AFO datasets, and ablation experiments examine the contribution of each component. On Person Detection in Water, HPSNet achieves 78.2% mAP@50, 37.5% mAP@50:95, and 67.2% recall, improving the YOLOv11 baseline by 2.7, 2.2, and 4.2 percentage points, respectively. On AFO, it achieves 88.4% mAP@50 and 59.4% mAP@50:95, with gains of 0.7 and 1.9 percentage points over the baseline. The model contains 4.18 M parameters, requires 9.6 GFLOPs, and processes an image in 13.4 ms on an RTX 4090. These results demonstrate improved detection of small and visually weak maritime targets relative to YOLOv11 while maintaining a moderate model scale, providing a foundation for future deployment and optimization on embedded maritime UAV platforms.