Roof Components of Metro Vehicles Detection Algorithm Based on Improved YOLOv5

Aiming at the problem of great labour intensity and high missed detection rate in manual detection of key components on roof of metro vehicles, an improved YOLOv5 target detection method is proposed. First, the lightweight design approaches are introduced, including the replacement of C3 module in Backbone network with Ghost_C3 module, and the substitution of Ghost convolution for the traditional 2D convolution, thus the lightweight YOLOv5 model is obtained to reduce the model complexity and computation consumption. Second, the CA attention mechanism is integrated into the Ghost_C3 module to strengthen the feature perception of roof target and improve the accuracy performance of the lightweight YOLOv5 model. The experimental results show that, compared with the original YOLOv5, the FPS of the improved YOLOv5 model reaches 102.04, the mAP is 97.98% which is improved by 1.36%, the model parameters are reduced to the original 33.98%, and the computation is reduced by 36.65%. The results also indicate that the proposed YOLOv5 can detect the key components on roof of metro vehicles effectively and can support the following research on the service condition analysis.