A deep helmet wearing detection method based on attention mechanism and stochastic weight averaging

Abstract

Helmet wearing is one of the most effective means of ensuring the personal safety of workers at job sites such as construction sites. Improving the accuracy of helmet wearing detection is one of the key technologies for intelligent helmet wearing supervision. To address the problem that the YOLO v5 target detection algorithm fails to focus on important features in the process of extracting features, a YOLO v5 algorithm based on the attention mechanism is proposed to pay attention to important features to improve the detection accuracy. Then, the model is optimized based on the idea of stochastic weight averaging to further improve the model detection performance. The specific method is as follows: After the training iteration until the model accuracy is stable, the learning rate is adjusted to train multiple model parameters, and the final weight model is obtained by stochastic weight averaging. The improved YOLO v5 target detection method has higher detection accuracy than Faster R-CNN, SSD, YOLO v3, YOLO v4, and other detection algorithms, with about 2.3% improvement over YOLO v5, which prove that attention mechanism and stochastic weight averaging are effective methods to improve the performance of helmet detection.