| Railway transportation is used for both freight and passenger transport,which is a major transportation artery related to the national economy and people’s lives.At the present stage,railway transportation is still unable to achieve fully enclosed operation,and railway traffic accidents often occur because of the weak railway infrastructure protection facilities.Pedestrian intrusion is one of the most important factors affecting railway traffic safety.Real-time pedestrian detection and alarm in key railway areas will provide guarantee for the train running safety.At present,short focus cameras are widely used in railway scene monitoring.In order to effectively solve the problems of high missed detection rate and high false detection rate caused by low detection accuracy of small-scale pedestrian in large field of view,this thesis studies the small-scale pedestrian detection algorithm for railway based on deep learning.The main work of this thesis are as follows:(1)Aiming at the problem of small-scale pedestrian detection in railway scene,a small-scale pedestrian detection algorithm based on attention and multi-level feature fusion was proposed.The basic pedestrian detection network was designed based on YOLOv3.Combined with small-scale pedestrian characteristics,the target detection network was improved and optimized from three aspects: multi-scale prediction,feature fusion guided by attention mechanism and loss function optimization.The experimental results show that the proposed algorithm has better detection performance than basic network,and the log average miss rate is reduced from 25.2474% to 9.1843%.(2)In order to make full use of the small-scale pedestrian motion information between video frames,a small-scale pedestrian detection network combined with motion information was proposed.The continuous video frames were input into the network as a whole,and the three-dimensional convolution was introduced to extract the pedestrian motion information between frames.At the same time,the adaptive feature fusion module was used to realize the adaptive aggregation of pedestrian spatial position information and motion information.The test results in the private dataset show that the proposed algorithm further reduces the log average miss rate from 9.1843% to6.8919%,and the false detection caused by partial occlusion and complex background interference is effectively improved.(3)To further reduce the missed detection rate of small-scale pedestrian,a small-scale pedestrian detection algorithm combined with multi-target tracking was designed.The multi-target tracking algorithm was used as the post-processing of the detection network to supplement the missing pedestrians on the track.The test results in railway scene show that the proposed algorithm can effectively improve the detection performance of small-scale pedestrian,and reduce the log average miss rate to 5.3527%.(4)In order to solve the problem of convolutional neural network with large amount of parameters and computation,the network structure was lightweight.A lightweight feature extraction network was proposed.The depthwise separable convolution was used to replace the conventional convolution to reduce the amount of network parameters and computation,and the inverted residual structure was introduced to improve the feature extraction ability of the model.The test results in railway scene show that the proposed algorithm can reduce the model size from 248.4MB to 110.9MB without significantly reducing the detection accuracy,and speed up the algorithm.This thesis is committed to improving the detection performance of small-scale pedestrian in railway scene.The network is optimized from four aspects: single frame input,continuous frame input,post-processing and lightweight.The experimental results show that the proposed algorithm can effectively reduce the missed detection rate and false detection rate of small-scale pedestrian,which has a good development prospect and practical application value. |
PDF Full Text Request