Research On Railway Foreign Body Intrusion Detection System Based On Deep Learning

The intrusion detection in the railway boundary has always been an important subject in the field of rail transit.Traditional frame difference detection algorithm gets disturbed by illumination changes and scene changes easily,and it is unable to meet the needs of railway on-line detection for a long time.Deep Learning has the characteristics of better expression and network training pattern,and shows great potential in the field of image,so bringing deep learning into Railway intrusion detection system is full of great significance.Main purpose of this paper is to design a detection system based on deep learning to achieve the automatic detection function of invasion objects(rock-fall,pedestrian,animal and car)and get an early warning.The intrusion detection system mainly includes target extraction,tracking,behavior analysis and recognition modules.(1)In the target extraction,because of the shortage of frame difference,the researchers use the Mixed Gaussian Model(GMM)and connected components labeling algorithm to realize the extraction of the foreground target.The experiment shows that the algorithm adapts to the actual environment.When multiple targets were detected,the researchers used the Nearest Neighborhood Data Association(NNDA)algorithm to correlate the same target in different frames,avoiding target loss and misalignment.In the tracking step,this paper studies mean-shift algorithm based on HOG,as mean-shift algorithm is effective to avoid the loss of the target due to the static target integration into the background,and HOG has a more reliable robustness than the color histogram features in the light changes and slight deformation of the object.(3)Behavior analysis and classification mainly based on the target behavior to determine the rules to make sure whether the alarm is allowed,and classification depends on ration of width and height.(4)According to the above algorithm,the railway intrusion detection system is showing that traditional features are not perfect enough to meet our requirements.Thus the researchers proposed DMSM(deep mean shift model)model.The traditional HOG or color histogram features are susceptible to interference such as background clutter,transformation,illumination change,occlusion,target cross motion,etc.,and are tested on a public library.The results show that DMSM has more good results,showing a strong potential and development prospects.