基于双源图像协同识别技术的驼峰抱闸车辆检测系统设计

Design of Detection System for Vehicles with Holding Brake in Hump Yard Based on Double-source Image Collaborative Recognition Technology

为有效识别驼峰编组站抱闸车辆,提高车辆溜放作业效率,设计研发基于双源图像协同识别技术的抱闸车辆检测系统。采用双摄像机分别检测溜放车辆闸瓦和气缸状态,以提高系统检测的准确率。以YOLOX算法为网络基本框架,设计搭建抱闸车辆目标检测模型;因车辆抱闸属于小概率事件,导致模型训练样本稀少,为此采用Mosaic和Mixup算法扩充数据集;结合抱闸车辆制动部件小目标运动特点,引入混合注意力机制和转置卷积方法进行特征增强,优化小目标检测能力。经现场试验验证,该系统满足驼峰溜放作业对实时性和可靠性的要求,有效提高复杂作业环境下抱闸车辆检测的准确性,以进一步保障编组站作业安全。

In order to effectively identify the vehicles with holding brake in the hump marshalling station and improve the efficiency of vehicles rolling operations,a detection system for vehicles with holding brake based on double-source image collaborative recognition technology is designed and developed.Dual cameras are used to detect the status of the brake blocks and cylinders of rolling vehicles separately to improve the accuracy of the system detection.Using the YOLOX algorithm as the basic framework of the network,this paper designs and builds a target detection model for vehicles with holding brake.Due to the training samples for the model are scarce caused by the low probability of vehicles with holding brake,Mosaic and Mixup algorithms are used to expand the datasets.Considering the motion characteristics of small targets in the braking components of vehicle with holding brake,this paper introduces a hybrid attention mechanism and transposed convolution method for feature enhancement to optimize the detection ability of small targets.Through on-site testing and verification,the system meets the requirements of real-time and reliability for hump rolling operations,effectively improving the accuracy of detection of vehicle with holding brake in complex working environments,and further ensuring the safety of marshalling station operations.