Research On Railway Foreign Object Intrusion Detection Algorithm Based On Scene Understanding

Object detection based on machine vision is the basic function of the railway foreign body intrusion monitoring system,which is important to railway operation safety.The accuracy of the existing intrusion object detection algorithms is low.To improve the object detection accuracy and make full use of the object size obtained from the structural information of the railway scene,the foreign intrusion object detection framework based on scene understanding is proposed.This thesis studies the railway scene structure information extraction algorithm,single view scene image spatial information recovery method,and known class and unknown class object detection algorithm that provides a new solution for railway foreign object intrusion detection.To obtain the scene structural information,the vanishing point and rail detection algorithm based on a deep multi-task convolution network is proposed.The proposed algorithm solves the problems of low accuracy of the traditional single task network in structural information detection.The multi-task network consists of a shared feature extraction base network and three sub-task networks,which are vanishing point regression,rail segmentation,and vanishing area segmentation.The two segmentation sub-tasks share the encoder-decoder structure and fuse multi-scale features in the decoder.Because of the assistance of rail and vanishing area segmentation tasks and multi-scale feature fusion,the multi-task network improves the detection accuracy of the vanishing point in different error ranges,enhances the generalization performance of the model,and saves computing resources.To solve the problem of automatic estimation of the spatial size of the railway intrusion object,a spatial information recovery method based on scene structure understanding is proposed.Aiming at the problem that traditional camera calibration algorithm needs manual intervention and offline operation,prior knowledge of the scene structure is used to establish the height information reconstruction model.The absolute mapping relationship between image height and spatial height is estimated through vanishing point and fixed railway equipment detection.The effectiveness of the proposed algorithm is verified in the railway scene.To reduce the missed detection rate and false detection rate of the existing deeplearning-based object detection method,a known class object detection method that integrates the scene structure information is proposed.Firstly,the vanishing point that contains perspective geometry information is used to estimate the relative depth of the scene.Then the distant area that small scale object centrally distributed is located and rescaled.The existing object detection network obtains the auxiliary input through the equalization of heterogeneous scales.The credibility of the detected object is further analyzed by using the recovered spatial information.The experimental results show that the equalization of heterogeneous scales improves the detection rate of existing object detection network in the distant area and spatial information reduces the false detection rate.To solve the problem of low detection accuracy caused by dynamic background elements,a foreground object detection algorithm based on adaptive segmentation threshold is proposed.The temporal dynamic of pixel intensity,feedback information of detection result,and spatial information of super-pixel is used to determine the dynamic adjustment factor of the threshold to follow scene change.The foreground pixels in the initialization sequence are eliminated by using the spatial-temporal information of the pixel to avoid the false detection caused by the existence of moving objects.The comparison experiments on railway scenes show that the proposed algorithm improves the comprehensive accuracy of foreground object detection.Based on the foreground detection,the dimension attribute of unknown foreign objects is obtained through scene understanding.Based on the above research results,the proposed framework realizes the reliable recognition of known class objects and obtains the size attribute of unknown class objects,which lays the foundation for further improving the performance of railway foreign object intrusion alarm system.