Research And Implementation Of Video Pedestrian Detection System Based Ondeep Learning

With the continuous advancement of computer vision technology,a large number of video surveillance systems have already existed in our lives.Whether in the fields of security or business,pedestrian detection and pedestrian counting based on computer vision are of great significance.In the actual monitoring scene,changes in the pedestrian's posture,pedestrian occlusion,and complex background conditions will all affect the final result.Researchers at home and abroad have conducted in-depth research in various fields such as target detection,target tracking,and pedestrian counting,but the use of fusion is relatively small.Therefore,this paper studies the above fields based on deep learning methods and uses pedestrian detection+pedestrian tracking+counting algorithms to implement a video pedestrian detection and counting system for scenes such as two-way pedestrian flow at the entrance of scenic spots.This paper uses the YOLOv3 algorithm as the basis for pedestrian detection,uses a lighter network MobileNetv2 to replace the backbone network in the original YOLOv3 for feature extraction,and integrates adaptive spatial feature fusion strategies into the network model structure to improve multi-scale detection.The K-Means clustering method was changed to the K-Median clustering method to obtain a new anchor box,GIOU was introduced to design a new loss function,and finally,a new data set was independently produced based on the public data set for model training and testing.The improved algorithm has improved detection accuracy and detection speed.Use the Deep Sort algorithm to achieve pedestrian tracking,introduce the generalized intersection ratio and replace the intersection ratio in the original algorithm to compare the detection frame and the prediction frame to achieve a better matching effect;in the cascade matching aspect,the problem of pedestrian occlusion is proposed.The method of increasing the standby time of lost frames enhances the accuracy of tracking and effectively reduces the problem of pedestrian ID jumps.Use the cross-line method to count pedestrians,and design each pedestrian ID serial number to record only one line touch,avoiding errors caused by pedestrians wandering back and forth near the counting line,and reducing the impact of pedestrian ID jumps in the tracking algorithm on the count.Finally,based on the above key technologies,this paper designs and implements a video pedestrian detection and counting system based on deep learning.