Lightweight Railway Intrusion Object Detection Using Onboard Forward-Viewing Camera

Railway environmental safety is one of the important conditions for train operation safety.The intrusion objects seriously threaten railway security and brings huge risks to national property and people’s security.Intrusion detection methods based on video surveillance play an important role in the monitoring of train operating environment.Existing video surveillance includes fixed-point surveillance on the ground,drone surveillance and train front-view surveillance.Compared with the former two,the train front-view monitoring has the advantages of full route coverage,less influence by weather,and low data transmission delay.Due to the fast train speed and less computing resources of on-board equipment,real-time and easy-to-deploy are put forward for the intrusion detection method based on on-board forward-viewing video.Therefore,from the perspective of lightweight,this paper develops an intrusion detection method based on forward-looking video.Specifically,it includes the lightweight research of dividing the track area into the railway warning area,object detection,and detecting whether the object is in the track area:(1)Aiming at the problem of low accuracy of existing real-time semantic segmentation methods,this paper proposes a method for railway track region segmentation based on improved Fast-SCNN,and designs a feature fusion module,which improves accuracy while ensuring real-time performance.Rate.The superiority of our method is demonstrated by comparing different semantic segmentation methods on the Rail Scapes dataset proposed in this paper.(2)Aiming at the problems of large model and slow running speed of the existing railway object detection method,this paper uses the Tensor RT framework to optimize the inference process of YOLOv5 s,which improves the speed of the model and reduces the memory usage.We deploy the optimized model to embedded platforms Jetson Xavier NX,Jetson Nano.Several sets of test experiments are carried out on the onboard forward-viewing intrusion detection dataset proposed in this paper,which proves that it can meet the real-time requirements on embedded devices.(3)Aiming at the problems of unreasonable design of orbit segmentation and target detection process in existing methods,which leads to waste of computing resources and high false negatives.This paper designs a railway intrusion detection pipeline that only runs the segmentation method when an object is detected.A pseudo-code for judging the location of objects by combining object detection and track area segmentation is given,which saves computing resources and reduces false negatives.The accuracy of the method proposed in this paper is 98.26%,the false alarm rate is less than 0.1%,the model size is 26.1Mb,and the inference speed is greater than 25 frames/second,on the Jetson Xavier NX embedded platform.To sum up,this paper takes the railway onboard forward-looking video as the research data,and conducts research from the lightweight track area segmentation and the intrusion target detection.The advantages of this method are high precision,realtime,and easy deployment.