Study On Small Scale Pedestrian Detection For Railway Video Surveillance And Deployment On The ARM Platform

As railway transportation is an important channel for the circulation of people and materials in China,ensuring the safety of traffic is the top priority of railway transportation.Accurate and timely detection of errant pedestrians can effectively avoid railway traffic accidents.Traditional human video surveillance technology cannot meet the needs of high-performance and real-time detection,and the railway security mode gradually migrates to intelligence.In this context,this thesis studies the small-scale pedestrian detection algorithm for railway video surveillance and combines it with the embedded ARM platform to deploy the algorithm to realize the intelligentization of railway security.The main work of this thesis are as follows:(1)Aiming at the problems of low detection accuracy and slow speed of traditional detection methods,a detection method based on deep learning is adopted to realize pedestrian detection.The basic pedestrian detection network is designed based on YOLOv5 s,the lightest of the YOLOv5 series networks,and the railway dataset is jointly used with the COCO public dataset as the network training set to enhance the model generalization capability.The experimental results show that by using the joint dataset to train the YOLOv5 s basic network,the log average miss rate of the railway test set is19.49%,which is better than the detection results using a single training set.(2)Aiming at the problem of small-scale pedestrian detection in railway scenes,the loss function is optimized and an attention mechanism-guided feature fusion algorithm is designed.The α-CIo U(Alpha-Complete Interdection over Union)loss function is used to improve the small-scale target bounding box regression accuracy,while the coordinate attention mechanism is introduced to suppress the background noise of low-level feature maps in feature fusion,so that the network focus more on learning the semantic information of small-scale pedestrians.The experimental results show that the designed algorithm has better detection performance than the basic network,and the log average miss rate is reduced from 19.49% to 10.70%.(3)The deployment of a small-scale pedestrian detection network model on the embedded platform is completed.A model conversion scheme of the Hisilicon NNIE(Neural Network Inference Engine)is designed for the small-scale pedestrian detection network model,and the pedestrian detection algorithm is deployed on the embedded Hisilicon Hi3516DV300 platform.Finally,the detection accuracy and processing speed of the pedestrian detection system are verified.The log average miss rate is 21.65% and the processing speed is 14.02 FPS on the hardware platform,indicating that the hardware deployment scheme in this thesis can effectively complete pedestrian detection in railway scenarios.The small-scale pedestrian detection algorithm in the railway scene and the deployment scheme of the ARM platform proposed in this thesis are tested with the actual railway scene datasets.The results show that the research results have good application prospects.