Research On Algorithm Of Detection And Recognition For Obstacle In Front Of The Locomotive Based On Binocular Vision

Railway transportation is an important infrastructure for cargo transportation and mass travel,so it is of great significance to ensure the safe operation of locomotives.China's railway industry has developed rapidly,and its operating mileage and road network density have increased significantly.However,China's railway covers a wide range of terrain,and the locomotive operating environment is complex.The locomotive operation safety accidents caused by foreign matter intrusion frequently occur,which seriously affected the safety of railway transportation.In view of the landslide and debris flow caused by natural disasters and human-induced construction waste,as well as pedestrians,vehicles and other infringements,only the existed monitoring measures such as driver's lookout,trackwalker protection,and crossing monitoring are insufficient to meet the whole process of driving safety monitoring.Aimed at the existing problems,this thesis proposed a research on obstacle detection and recognition in front of locomotive operation based on binocular vision.The target sequence frame is analyzed,the rail is extracted as the bounding reference object,the suspected obstacles in the limit are located,and the deep learning-based detection and recognition algorithm is used to detect and identify the suspected obstacles.The main research contents of this thesis include:(1)Construction of the railway clearance.For the railway clearance model,the problem can not be established accurately in real time.Firstly,the image acquired by the visual sensor is preprocessed.The rail in the image is extracted according to the gauge invariant characteristic as the basis for establishing the limit.The segmented curve model is established to divide the rail into the near view and the distant view.In the near-field area,the multi-constraint Hough transform is used to obtain the straight-track model.In the distant view area,the feature points are searched into the Catmull-Rom curve function to fit the curved track model,and the bounded range is obtained according to the proportional expansion of the extracted rail curve.(2)Suspected obstacle target location.For the target interference in the target location of suspected obstacles,the disparity map is initialized by pre-processing the binocular image sequence frame,the V-disparity map is established according to the disparity map,and the suspected obstacles are extracted in the image in the V-disparity map.The coordinate position records and transmits the suspected obstacle coordinate position to the suspected obstacle detection and recognition module.(3)Suspected obstacle detection and identification.Aimed at the problem of low real-time detection in the detection and identification of suspected obstacles,the YOLO algorithm based on deep learning is used to extract the convolution features of the suspected obstacles in the image and compress the input feature map through the pooling layer tocompress all the features.The full connection layer output prediction frame is used to detect and identify the suspected obstacles,and the target position of the suspected obstacles in the previous module is compared with the output prediction frame to judge the output recognition result.According to the experimental results in different environments,the proposed algorithm for detecting obstacles in front of locomotives based on binocular vision had higher recognition accuracy and better real-time performance.