Research And Implementation Of Track Abnormal Object Detection Based On Multi-sensor Information Fusion

At present,in the detection system of foreign matter intrusion in rail,the traditional passive detection has problems such as low detection accuracy,poor track foreign matter category and position positioning effect,and there will be hidden dangers such as tool loss and missing detection in the regular track inspection operation,which will appear in the actual track safety inspection system.In recent years,researchers have combined video image analysis technology and artificial intelligence technology to apply rail foreign matter detection which can monitor the train operating environment and operation safety in real time.However,under the conditions of poor lighting conditions or rainy and foggy weather,the single-sensor-based rail foreign matter detection algorithm will have a series of problems on the embedded development platform of practical applications,such as low detection accuracy,slow detection speed,and high model deployment cost.In this thesis,the rail foreign matter detection algorithm is lightweighted and a variety of sensors are used for image data acquisition,and the detection results of multi-sensor image information data are fused,so as to improve the accuracy of rail foreign matter detection,reduce the probability of false detection and missed detection,and improve the safety of train operation.Secondly,the lightweight operation can also reduce the size of the algorithm model,thereby facilitating the deployment of the model and improving the speed of rail foreign matter detection.The specific research content is as follows:(1)In the thesis have made an improvement to the original YOLOv5 algorithm,which cannot meet the real-time detection in the actual engineering project process.In this thesis,by lightening the operation,the backbone network in the original YOLOv5 algorithm is replaced by lightweight operation,so that the lightweight detection algorithm can be deployed on the actual embedded development platform to achieve a real-time detection effect.(2)Although the lightweight rail foreign matter detection algorithm can achieve a real-time detection effect in terms of detection speed,but there will be a certain decrease in detection accuracy,this thesis uses a new pooling module and a new loss function for the lightweight detection algorithm,so that the detection accuracy of the lightweight algorithm can be improved to ensure that the algorithm does not lose detection accuracy while lightweighting.(3)This thesis uses a variety of sensors for image data acquisition,including three sensors:lidar,telephoto camera,and close-focus camera.Different sensors are suitable for different practical scenarios,and the image data of each sensor can complement each other,and through the fusion of multi-sensor image information data detection results,the lightweight rail foreign matter detection algorithm improved in this thesis can be applied to more complex practical scenarios to improve the safety of rail train operation.