Research On Railway Pedestrian Invasion Limit Detection System Based On Deep Learning

In recent years,China’s rail transportation industry has developed rapidly.Passenger traffic,rail freight volume,operating speed and mileage have been greatly improved,and driving safety has received more and more attention.Intelligent analysis of foreign objects in front of traffic based on image intelligence has become one of the research hotspots in the safety monitoring of railway perimeter.However,most of the existing intrusion detection systems are based on fixed camera collection,which has the disadvantages of low coverage and difficult information exchange.To this end,this paper proposes an image processing-based railway pedestrian intrusion detection algorithm,the main contents are as follows:(1)The track detection algorithm under low-illuminance conditions is proposed,which realizes the uninterrupted track detection of the railway locomotive for the whole period.First,analyze the characteristics of low-illumination railway environment images such as evening and night,invert such images,and adopt the dark channel prior enhancement algorithm for image enhancement to ensure the integrity of pedestrian features in low-illuminance images.Then,analyze the characteristics of the rail image,combine the edge information,LUV,HLS and other features to detect the rail,optimize the initial position of the rail fitting and the fitting function,and realize the extreme "S" rail detection.The results show that the rail detection algorithm can be applied Most of the railroad scenes.(2)Proposed a fusion algorithm of visible light and thermal imaging images suitable for railway scenes.Considering that the image of the thermal imaging camera can clearly reflect the edge of temperature change,and the texture and detail characteristics of the image of the visible light camera are clear,the advantages of the two are combined.The edge detection of the thermal imaging map is used to detect The method of superimposing the obtained edge information onto the visible light image combines the advantages of the two and merges to generate a new image.Combining the clear advantage of pedestrian edge information with thermal imaging and the obvious advantage of texture color features of visible light images,the results show that this method can improve the characteristics of pedestrians in a single image.(3)A new lightweight mobile pedestrian detection algorithm Mobile-YOLO for embedded devices is proposed.The new algorithm is based on the target detection neural network YOLOv3.Feature extraction uses a lightweight network model MobileNetv2.The model anchor clustering algorithm improves the initial point selection method.The detection layer adds a new receptive field and optimizes the algorithm.The YOLOv3 model is optimized for a railway pedestrian detection algorithm that can be calculated in real time on low-power embedded devices,and is trained based on the established railway pedestrian data set.Experiments show that the new algorithm basically guarantees the accuracy of the original algorithm for pedestrian detection while greatly improving the current With the real-time nature of the algorithm,it can better meet the needs of actual railway high-speed detection under the existing hardware configuration conditions.(4)A complete software and hardware system for railway pedestrian intrusion detection is designed.According to the detection requirements,a hardware system consisting of a telephoto visible light camera,an infrared thermal imager,an NVIDIA Jetson TX2 processor and a touch screen was determined,and a software system including three subsystems of image enhancement,rail detection,and pedestrian detection was realized,based on Python and PyQt Designed a host computer interface suitable for the railway industry.The system was installed and debugged on the DF7 locomotive,and its feasibility was initially verified.