基于改进YOLOV5m的列车风管与折角塞门检测

Detecting Duct Picking and Angle Cock of Train Based on Improved YOLOV5m

由于操作环境复杂,风管和折角塞门尺寸小,原始YOLOV5m算法的检测精度有待提高。为此,在YOLOV5m的检测头输出部分增加了一个大尺寸特征图,用于提升小尺寸目标的检测能力。首先在主干网络中增加1个含有多卷积、短连接的类残差分流模块,以丰富浅层网络的梯度流信息,并将新增模块的输出作为新增的大尺寸特征图。然后在Neck部分使用特征金字塔模型FPN和路径聚合模型PAN进行多尺寸特征融合,最终在检测头部分得到了4种不同尺寸的特征图。结果显示,改进YOLOV5m的检测帧率约为每秒100帧,比原始YOLOV5m在平均精度mAP上提升了1.3%,表明改进YOLOV5m可作为可见光辅助模型提升列车摘钩机器人多模态识别系统的性能。

Due to the complexity of the operating environment and the small size of the duct pickings and angle cocks,the detection accuracy of the original YOLOV5m algorithm needs to be improved.To this end,a large-sized feature map has been added to the detection head output section of YOLOV5m to enhance the detection capability of small-sized targets.Firstly,a class residual diversion module with multiple convolutions and short connections is added to the backbone network to enrich the gradient flow information of the shallow network,and the output of the new module is used as a newly added large-sized feature map.Then,in the Neck section,the feature pyramid model FPN and path aggregation model PAN were used for multi-scale feature fusion.Finally,four different sized feature maps were obtained in the detection head section.The results show that the improved YOLOV5m has a detection frame rate of about 100 frames per second,which is 1.3%higher than the original YOLOV5m in terms of average accuracy mAP.This indicates that the improved YOLOV5m can be used as a visible light assisted model to enhance the performance of the multi-modal recognition system for train uncoupling robots.