Real-Time Helmetless Detection System for Lift Truck Operators Based on Improved YOLOv5s

Abstract

Safety helmet plays a major role in protecting the safety of operators in industry, and several helmetless detection methods have been developed based on artificial intelligence. However, existing detection methods cannot work well for machinery operators in specific scenarios, such as factory environments, in which low accuracy and efficiency for small helmet targets could happen occasionally. Aiming to comprehensively handle these issues, this paper proposes a real-time helmetless detection method for lift truck operators based on improved YOLOv5s. Firstly, the lightweight multiscale attention EfficientViT is added to improve the detection accuracy for small-sized helmets. Secondly, the detector C2 F-Net from Transformer structure is added to improve the predictability of challenging occurrences. In addition, the loss function is changed to Alpha-IoU, which further enhances the detection ability of small-sized targets. Finally, a real-time helmetless detection system is built with a set of well-designed detecting logic. The system effectively implements the proposed method and provides real-time monitoring and detection of helmetless lift truck operators. By conducting experiments on a self-created dataset derived from factory surveillance videos, this paper successfully validated the effectiveness of the proposed method. Specifically, the results showed that the proposed method increases the mAP (0.5) of the original algorithm in abnormal class by 6.7%, reaching 98.7%, and the mAP (0.5:0.95) is improved by 6.5%, indicating the proposed method shows significant improvement in real-time detection performance for operators not wearing safety helmets in specific scenarios.

Keywords

• YOLOv5
• helmetless detection
• EfficientViT
• C²F-Net
• Alpha-IoU