Pedestrian Detection Of Parking Based On YOLO And Improved KCF

Parking lot as an environment that people often come into contact with in their daily life,the management of parking lot has always been a difficult problem because of the influence of various factors.In recent years,with the rapid development of China's economy,the population continues to expand,and the car ownership and traffic volume are also on the rise.However,the supply of parking facilities in the city is seriously insufficient,which leads to the increasingly prominent contradiction between the supply and demand of parking facilities.How to do the effectively abnormal behavior detection in the parking lot,so as to improve the management efficiency of the parking lot and reduce the management cost arises at the historic moment.Therefore,this paper combines the YOLOv3-Tiny algorithm and the improved KCF algorithm,and puts forward the judgment of the static target after the movement,and finally realizes the detection and recognition of five kinds of abnormal behaviors in the parking lot,which are not parking according to the regulations,occupying the parking space illegally,loitering in the parking lot,staying by the side of the pedestrian for a long time,and potential violence in the parking lot.Firstly,according to the specific situation of this paper,we collected a data set of 500 pictures of five kinds of targets: club,mask,knife,gun and car.Then we combined the INRIA pedestrian data set to form a new data set.In order to improve the training effect,YOLOv3-Tiny can learn more details.In this paper,the dark channel defogging algorithm with fast guided filter is used to optimize the dataset images.At the same time,in order to make better use of computing resources and give full play to the advantages of YOLOv3-Tiny,K-means clustering algorithm is used to obtain the optimal number and dimensions of anchor boxes according to the width and height of different targets in the data set,and the optimal number is selected by comparing the performance of different anchor boxes.Thus,the target in the screen is effectively detected and recognized,and whether there is potential violence in the parking lot is directly determined from the armed and masked behaviors.Secondly,in order to determine three kinds of abnormal behaviors,which are not parking according to the regulations,occupying the parking space illegally,staying by the side of the pedestrian for a long time this paper proposes a static target detection algorithm based on improved frame difference algorithm and vibe algorithm.It is used to detect the static target object after entering the screen.In the ROI region set in the video frame,the algorithm is used to determine whether it is a still foreground target,and then it is matched with the recognition target of YOLOv3-Tiny to determine whether it is the above two abnormal behaviors.Then calculate the Euclidean distance between the stationary pedestrian and the nearby stationary vehicle to determine whether it is a long-time vehicle side stop.Finally,this paper studies the KCF(Kernelized Correlation Filters)algorithm,aiming at the performance that the KCF algorithm will cause tracking error and drift in response to the fast moving target,this paper proposes a tracking error judgment mechanism based on multiple features,for the target judged as tracking error.In this paper,the multi-sampling KFC target tracking algorithm is used to multiple-sample blocks of the target with tracking error so as to retry it.Finally,describe the movement track of pedestrians,and determine whether they are wandering in the parking lot according to the movement track of pedestrians.The experimental results show that the algorithm can effectively detect the five abnormal behaviors and meet the real-time requirements.