Design And Implementation Of Lightweight Pedestrian Detection Network For Railway Scene

The increasing railway mileage puts forward higher requirements for railway security.Traditional railway security relies on human video surveillance to monitor intruders.Now this kind of traditional security mode should also move forward to intelligence.This paper combines advanced target detection technology and embedded heterogeneous processor to build an embedded video pedestrian detection system suitable for railway scene.The main tasks of this paper are as follows:(1)The railway scene image data set is divided into training set,test set and supplementary fine-tuning data set.The general target detection network YOLOv3 is utilized to design a special pedestrian detection network.The network is optimized according to the characteristics of railway scene.Average precision and log-average miss rate are introduced to comprehensively evaluate the network detection accuracy.(2)The method of channel pruning is used to realize the lightweight of YOLOv3 pedestrian detection network.In this paper,the pruning strategy of regular channel is designed to improve the hardware accelerator friendliness of the limit pruning strategy.The supplementary fine-tuning data set is introduced to improve the problem that the detection accuracy cannot be recovered sufficiently after fine-tuning on the original train set,so that the detection accuracy is higher than that before pruning when the network is compressed on a large scale.(3)The lightweight pedestrian detection network is realized on the embedded heterogeneous device ZCU104,which promotes the practical application of the pedestrian detection network.Input and output double buffered image queues and multithreaded scheduling are designed to optimize video streaming.The accuracy and processing speed of the pedestrian detection system are verified,and the expected design goal is achieved.This paper has completed the data set arrangement,algorithm design,network pruning,and the lightweight realization on the special system.After testing,the lightweight pedestrian detection system designed in this paper can preliminarily meet the detection requirements in railway scenes.