基于图像识别的弓网接触点检测方法

Pantograph-catenary contact point detection method based on image recognition

针对现有受电弓-接触网(弓网)接触点检测方法无法兼顾实时性与准确性的问题,提出两阶段快速检测方法.在第1阶段提出基于改进BiSeNet v2的弓网区域分割算法.采用浅层特征共享机制将细节分支提取的浅层特征送入语义分支中获取高层语义信息,减少冗余参数;将压缩激励注意力模块嵌入网络中,增强重要通道信息;加入金字塔池化模块提取多尺度特征,提高模型精度.在第2阶段,基于分割结果,使用直线拟合和位置校正实现接触点的检测.实验结果表明,所提分割算法精度为87.50%,浮点运算数为6.73 G,在CPU(Intel Core I9-12900)和JETSON TX2上推理速度分别为49.80、12.60帧/s.所提检测方法在弓网仿真平台和双源智能重卡的弓网系统中进行实验,实验结果表明,该方法能够有效检测弓网接触点.

A two-stage fast detection method was proposed aiming at the poor real-time performance and low accuracy of existing pantograph-catenary contact points detection methods.In the first stage,a pantograph-catenary region segmentation algorithm was proposed based on the improved BiSeNet v2.The shallow feature sharing mechanism was used to send the shallow features extracted from the detail branch to the semantic branch to obtain the high-level semantic information and reduce the redundant parameters.The Squeeze-and-Excitation Attention Mechanism was embedded into the network model to enhance the important channel information.The Pyramid Pooling Module was added to obtain the multi-scale features to improve the accuracy of the model.In the second stage,based on the segmentation results,contact points detection was achieved by the linear fitting and the position correction.The experimental results showed that the proposed segmentation algorithm had an accuracy of 87.50%,floating point operations of 6.73 G,and an inference speed of 49.80 frames per second and 12.60 frames per second on CPU(Intel Core I9-12900)and JETSON TX2.The proposed detection method was experimented in the pantograph-catenary simulation platform and the pantograph-catenary system of the dual-source intelligent heavy truck.The experimental results showed that the method can effectively detect the contact points of the pantographcatenary.