Research On Technologies Of Image Recognition For Railway Operation Safety Based On Deep-Learning

With the rapid development of our country’s railways system,the scale of railway network continues to expand,the image and video collection points related to railway operation safety have increased,and big data on railway operation safety continues to grow.At present,the image analysis and fault detection for these data are mostly done manually.Such problems as low detection efficiency,high detection costs,false detection,and missed detection are existed.Therefore,to improve the intelligent analysis,recognition and detection capabilities of structured data such as railway operation safety videos and images,as well as to enhance the warning and prediction capabilities of railway operation safety detection and monitoring systems.There is an urgent need to carry out research on key technologies of intelligent image recognition for railway operation safety,which provide a technical support to ensure safe and stable operation of railways.Over the past few years,various algorithms based on deep learning have been widely utilize in face recognition,detection and driverless vehicles due to their deep network structure and strong nonlinear mapping ability.However,these methods’ application in railway traffic safety is relatively rare.This thesis uses the object detection method based on deep learning in railway traffic safety,and studies the deep neural network based detection of locomotive signal lights,pedestrians on railway tracks;the detection of open status and open direction of railway turnouts;the recognition of signal lights,pedestrians,and locomotives in railway shunting operations.The findings provides support for intelligent analysis and decision-making of railway operation safety,and propose a reliable guarantee for eliminating hidden dangers of railway operation safety.This thesis has carried out the following works:(1)In order to address the detection problem of railway signal lights and pedestrians on railway tracks during the moving process of locomotives,a YOLOv4 based recognition method of locomotive signal lights and pedestrians on railway tracks is proposed,and an intelligent auxiliary lookout system for drivers is designed and developed based on the proposed method.Moreover,a recognition dataset of locomotive signal lights and pedestrians on railway tracks(LSLPRD)is established.The experimental results on LSLPRD dataset show that,our method is able to detect the locomotive signal lights and pedestrians on railway tracks in the moving process of locomotives with high recognition precision,the detection accuracy reaches 84.75%,and the detection speed achieves 25 FPS.The intelligent auxiliary lookout system is applied in the locomotive depot of Lanzhou Railway Bureau,the results reveal that the system is able to meet the real-time detection requirements of signal lights and pedestrians within the detection range of 1.2km under different illumination conditions.(2)To meet the recognition requirements of signal lights,pedestrians and rolling stock in the process of railway shunting operation,a lightweight object detection method of railway shunting operation based on improved YOLOv4-tiny is presented,and a comprehensive recognition dataset of railway signal lights,pedestrians,and rolling stock(LSLPRD+)is established.Our method greatly improves the detection ability and recognition accuracy of small objects without reducing detection speed by introducing the multi-scale feature fusion module and SE-Net attention module into the backbone network of YOLOv4-tiny.The experimental results on LSLPRD+ dataset show that,the detection speed of our method on NVIDIA Jetson Nano reaches 21 FPS,and the detection accuracy reaches 96.04%.Compared to YOLOv4-tiny,which is also a lightweight object detection method,the proposed method has better detection ability for small objects and distant blurred objects,and effectively reduces the missed detection rate of small objects.(3)In order to resolve the problem of recognition and detection of railway turnouts,a Mask R-CNN based detection method of railway turnouts open state and open direction is proposed,and a railway turnout detection dataset(RTRD)is established.Our method improves the recognition and segmentation speed of railway turnouts by using railway switch template matching method to extract the key frames of video.The experimental results on the RTRD dataset show that,in the detection range of 5m～1200m,the recognition accuracy of the proposed method reaches 83.5%,and the detection speed achieves 10 FPS.Compared with Faster R-CNN and YOLOv4,The method can accurately detect the driving state and driving direction of different turnouts,and realize the detection of the direction of the turnout tip rail,and the detection accuracy is increased by 2.5% and 9.5% respectively.(4)To address the detection problem of railway tracks and turnouts in the process of railway shunting operation,a segmentation and detection method of tracks and turnouts based on improved YOLACT is proposed,and a comprehensive dataset of railway tracks and turnouts detection(RTRD+)is established.Our method significantly improves the segmentation and detection ability of tracks and turnouts by introducing the deformable convolution and CBAM attention module.The experimental results on RTRD+ dataset reveal that,compared to Mask R-CNN and YOLACT,the proposed method is able to accurately detect and segment different tracks and turnouts areas,the segmentation accuracy is 4.05% higher than that of YOLACT,reaching 86.51%.The detection speed of our method on Ge Force GTX 2080 reaches 31 FPS,which is 23 FPS and 9 FPS higher than that of Mask R-CNN and YOLACT,respectively.Besides,the detection speed on NVIDIA Jetson Nano reaches 20 FPS.Findings of the key technologies of intelligent image recognition of railway safety enriches the research content of railway running and shunting safety,and propose references for the research on the realization of automatic locomotive driving and environmental perception technology.