Safety helmet detection method for underground coal mine scenarios based on improved YOLOv10

Mine safety accidents are closely linked to unsafe behaviors of coal mine workers, particularly the failure to wear safety helmets. Traditional manual monitoring methods, relying on video surveillance, are time-consuming, prone to errors, and unable to provide continuous monitoring. With the widespread use of surveillance cameras in mines, computer vision techniques have emerged as an effective solution for detecting unsafe behaviors. Existing detection methods, however, often struggle with the challenging underground environment, including low light and coal dust, which degrade detection accuracy. Additionally, these models tend to have high computational costs and long processing times, hindering realtime alerts. This paper proposes a novel safety helmet detection method for underground coal mines based on an improved YOLOv10 model (YOLOv10-MSH). The model integrates contrastive learning for data augmentation, enhancing sample diversity and robustness. It also introduces a Hybrid Attention Mechanism, combining spatial and channel attention to improve feature extraction and boost detection accuracy. Experimental results show that YOLOv10-MSH significantly improves both detection accuracy and speed, effectively meeting the requirements for real-time safety helmet detection in underground mining operations.