Toward Large-Scale Non-Motorized Vehicle Helmet Wearing Detection: A New Benchmark and Beyond

Abstract

The challenge of detecting helmet-wearing on non-motorized vehicles within road traffic scenarios has long been beset by issues like inadequate feature extraction and background noise interference. To address these challenges, an algorithm tailored for detecting helmet-wearing on non-motorized vehicles amidst complex road traffic environments was proposed in this paper. This algorithm employs feature enhancement techniques and context-aware fusion strategies to effectively address the considerable challenges presented by the vast quantity of non-motorized vehicles, small target dimensions, and the need for accurate violation monitoring in real-world settings. Specifically, the algorithm integrates a hybrid heuristic attention mechanism to refine local feature extraction capabilities. By amalgamating global and local features, it enhances the detection and recognition capabilities for diminutive targets. Furthermore, with the adaptive Transformer module addressing the challenges of target localization and occlusion in dense scenes, we propose an object detection network YOLO-HD for Non-Motorized Vehicle Helmet Wearing Detection. Moreover, to mitigate the issue of scarce data in real-world non-motorized vehicle datasets, a large dataset called NVHD-20K is carefully created to detect non-motorized bicycle helmets. A novel annotation methodology is employed to discern between stationary and moving non-motorized vehicles, thereby reducing false positives. Experimental results substantiate the efficacy of the YOLO-HD, attaining a commendable detection accuracy of 94.2% for diminutive targets like helmets. This surpasses the performance of contemporary state-of-the-art algorithms, thus underscoring its significant practical utility.