基于深度学习的轨道表面异物识别方法

Foreign Object Recognition Method for Track Surface Based on Deep Learning

针对现有异物识别方法存在识别精度低、成本高的问题,提出一种基于深度学习的视觉感知方法,对车辆在轨道表面上垂直投影区域的异物进行识别。首先,采用编码与解码框架构建车辆在轨道表面上垂直投影区域的语义分割模型,利用轻量级主干网络MobileNet v3作为特征编码器,并基于LR-ASPP解码器实现特征融合以提升分割精度;其次,从2个方面对YOLOX-s检测模型进行优化,利用深度可分离卷积作为残差单元的特征提取算子,并充分利用算子的高效性降低模型的复杂度,同时在尽可能不影响模型推理耗时的情况下,嵌入通道注意力机制以对特征进行加权处理;最后,基于自建图像数据集对识别方法进行精度验证。结果表明:所提语义分割模型平均交并比(Mean Intersection over Union,MIoU)达到了91.4%,单次推理耗时约为8.7 ms,能够有效地分割检测区域;所提检测模型平均精度(mean Average Precision,mAP)达到了81.07%,单次推理耗时约为10.8 ms。所提方法是深度学习技术在异物侵限方面的有效探索,可应用于多种实际场景,如检测入侵异物和划定危险区域等,对于提高铁路安全和效率具有现实意义。

To address the problems of low recognition accuracy and high cost of existing foreign object recognition methods,a visual perception method based on deep learning is proposed to recognize foreign objects in the vertical projection region of the vehicle on track surface.First,the vehicle’s semantic segmentation model on the track surface in the vertical projection region is constructed using the encoding and decoding frameworks;the lightweight backbone network MobileNet v3 is used as the feature encoder;and feature fusion is realized based on the LR-ASPP decoder to improve the segmentation accuracy.Second,the YOLOX-s detection model is optimized in 2 ways,using the deepwise separable convolution as the feature extraction operator of the residual unit and making full use of the efficiency of the operator to reduce the complexity of the model,while embedding the channel attention mechanism to weight the features with minimum impact on inference elapsed time of the model.Finally,the accuracy of the recognition method is verified based on the self-built image dataset.The results show that the proposed semantic segmentation model achieves the(Mean Intersection over Union)MIoU of 91.4%,and the single inference time is about 8.7 ms,which can effectively segment the detection region;the proposed detection model achieves the(mean Average Precision)mAP of 81.07%,and the single inference time is about 10.8 ms.The proposed method is an effective exploration of deep learning technology in foreign body intrusion,which can be applied to a variety of practical scenarios,such as detecting intruding foreign objects and delineating dangerous regions,and is of practical significance for improving railroad safety and efficiency.