LSTD-YOLO: A Lightweight Crowd Density Detection Method for Park Environmental Carrying Capacity Monitoring and Early Warning

With the rise of digital twins and smart cities, intelligent park management increasingly demands accurate real-time monitoring of visitor density to support spatial planning and policy decisions. Traditional methods primarily focus on total visitor counts, limiting their ability to capture instantaneous density, particularly in scenarios with uneven population distribution. Although unmanned aerial vehicle (UAV)-based monitoring has attracted attention for its mobility and wide field of view, it has struggled with small-target detection and limited onboard computational resources. To address these challenges, a lightweight object detection model named lightweight small target detection YOLO (LSTD-YOLO) was proposed for identifying small-scale pedestrians in UAV imagery. The model was enhanced with a small target feature fusion module and an additional detection layer to improve fine-grained feature extraction. In addition, the scale-based dynamic loss function was adopted to provide better supervision for small objects. Structured pruning and knowledge distillation were further employed to balance accuracy and efficiency, significantly reducing model complexity with minimal performance degradation. Experiments conducted on the CrowdDensity dataset demonstrate that the pruned LSTD-YOLO outperformed YOLOv11s, with the F1 score improved by 11.2%, mAP50 by 11.71%, and mAP50-95 by 5.32%, accompanied by only a modest 30% increase in GFLOPs. The proposed model also surpasses several state-of-the-art lightweight detectors. To facilitate future research, a high-quality UAV dataset named CrowdDensity has been released, comprising 41 684 annotated small pedestrian instances collected from nine urban parks in Nanjing.