Research On Pedestrian Detection Under Monitoring Scene

In recent years,due to the rapid promotion of deep learning technology,the field of object detection has been vigorously developed.Pedestrian detection technology has always been a hotspot in the research and application of object detection.For example,the development of technologies such as driverless,intelligent transportation and intelligent video surveillance cannot be separated from pedestrian detection.Among them,the intelligent monitoring system has been widely used in modern life.The main application environment of the pedestrian detection method studied in this thesis is for the scene under monitoring and it is an important part of the intelligent monitoring system.It can detect the pedestrians appearing in the surveillance video in real time,which has certain practical significance.This thesis first introduces the application background of pedestrian detection technology in the monitoring scene,and analyzes the development history and research status of the technology at home and abroad;Then several key technologies in the pedestrian detection method and the deep learning technology which is currently at the forefront of research are summarized.The object detection algorithm based on deep learning is applied to the pedestrian detection task to compare the detection performance of different models,and the optimal algorithm is selected;Finally,aiming at the problem that the pedestrian body is easily occluded when the people is crowded in the monitoring scene,this thesis proposes a detection model based on pedestrian head.At the same time,in order to meet the real-time detection requirements in the monitoring scene,the YOLOv3 that object detection algorithm of end-to-end is used to train the pedestrian head model.Considering that the YOLOv3 algorithm lacks target specificity,it is not fully applicable to pedestrian detection tasks in surveillance scenarios.So this thesis makes some improvements to this problem,the main work is as follows:(1)preparing the pedestrian data set,and marking the pedestrian head in the training set to establish the pedestrian head model;(2)focusing on the anchor box for the characteristics of the pedestrian head model,re-select the more suitable a anchor box number and dimension,and adjust the network structure;(3)model training and testing,parameter tuning,and get the final model.In order to verify the accuracy of the pedestrian head detection model in this thesis,the pedestrian image and surveillance video are used as the detection unit,and the model is tested by datasets such as INRIAPerson and PASCAL VOC,and the detection accuracy,missed rate and detection rate are calculated.At the same time,the model is compared with the pedestrian detection model based on Faster R-CNN,YOLOv2 and YOLOv3.The results show that the comprehensive performance in accuracy,missed detection rate and detection rate is better than other models.Experiments show that the work of this thesis is effective for implementing pedestrian detection in surveillance scenarios.