Fine-grained Detection Of Railway Track Pedestrian Based On Deep Learning

The development of high-speed railway has brought great convenience.The existing railway safety methods mainly rely on the combination of video surveillance and manual inspection.Because of the complicated and complicated railway track in China,the high speed of the train,and the staff lacking of professionalism,improper operation,negligence.Safety accidents still occur in railway operations.There is often an uninformed non-staff lack of safety awareness,ignore the warning signs,accidentally break into the railway areas,prone to serious accidents.However,due to the large scale of railway network in our country,it is difficult to determine whether there are illegal intrusions and stays on all routes by manual detection.Therefore,in order to improve railway safety,reliability and save labor costs,rail video surveillance system urgently needs comprehensive technical upgrading and large-scale application.However,the traditional algorithms often have some problems such as poor real-time and low precision.In addition,there are no published data sets for experiments in the industry.If the collected data is not balanced,it will have a great impact on the experimental process.In recent years,with the rise of deep learning,convolutional neural network algorithms acting on the field of vision have emerged one after another,and have achieved good results.Therefore,in view of the above tasks and difficulties,a fine grained pedestrian detection model is proposed in this paper.Based on YOLO V2 high performance detection algorithm,a hybrid attention mechanism is introduced in this method,which makes the algorithm perform well in a particular scene.The main work of this paper is as follows: a large scale railway pedestrian data set is constructed;Compared RCNN,Fast RCNN,Faster RCNN,YOLO and other deep learning algorithms,The paper decides to use the improved YOLOv2 model to achieve the target detection task.And the improved IOU algorithm is used to recluster anchor box,NMS was used to screen the candidate;A hybrid attention mechanism is introduced to meet the needs of complex scenarios and to improve the speed and accuracy of the detection system.Finally,the depth feature and the bottom feature of the image are fused to classify the pedestrian with fine-grained detection,Enable the system to distinguish between staff and non-staff while detecting travellers.