Improving the key point detection algorithm for railway operations based on YOLOv7-tiny

In recent years, deep learning algorithms have been widely applied across many industries. It has also seen rapid development in the field of key point detection. This paper proposes an improved YOLOv7-tiny algorith<strong>m f</strong>or detecting key points in railway workers’ operations, aiming to assist workers and reduce errors in their tasks. To enhance the model's detection capabilities, three optimization methods are proposed. Deformable Convolution (DCNv2) is introduced into the feature fusion module to enhance the model's ability to extract and fuse features. Next, to address the issues of object classification and localization, a lightweight decoupled head is introduced. Finally, to improve the accuracy of detecting targets with significant scale differences and enhance the model's convergence speed, we introduce SIoU as the loss function. Experimental results show that the improved algorithm achieves an accuracy (mAP&commat;0.5) of 91.1&percnt; and a detection precision (P) of 87.3&percnt, representing increases of 1.5&percnt and 4.8&percnt, respectively, compared to the YOLOv7-tiny model. The detection rate is 132 FPS, which meets the requirements for practical detection.