Pedestrian Tracking Technology And Implementation In Public Places For Surveillance Video

With the comprehensive deployment of video surveillance system in various scenes of people's daily life,the intelligent surveillance system which analyze video surveillance data through computer vision is increasingly used in traffic violation,prevention and control at people gathering region and other aspects,not only making the quality and efficiency of supervision and law enforcement at the basic level have been intensified,but also saving labor costs.As an important part of the system,the target tracking associates the same target in consecutive video frames to obtain the target's motion trajectory in the video,so it can analysis target's motion according to this trajectory.These video data are real-time video stream,it requires the algorithm to have a higher rate of processing.The method,which can make the algorithm run more faster and the calculated results more accurate by predicting the target location,has become more and more popular in the field of target tracking.Based on the above description,this thesis,looks on surveillance video of public places,predicts pedestrian motion based on their movement characteristics in the frame image and uses ways of pedestrian motion models and estimation of motion state,then combine with online target tracking to get pedestrians' motion trajectories.The following are what does this thesis research:1.Aiming at the problem of surveillance video frame dropping and low frame rate,through the research on video frame interpolation algorithms,the Super Slo Mo algorithm,which can interpolate frames in any time step,is selected to interpolate frames in the video.To make the original video have more frames,higher quality and become more fluency when it plays.2.In response to pedestrians should be detected in real time when following them,the thesis choose the YOLOV3 detection algorithm through the study of target detection algorithm;In response to Kalman filter's deficiency in recognizing how do pedestrians move,through the study of pedestrian movement model,the thesis choose the social force model because it is more suitable for pedestrians in public places;through the research on target tracking algorithms,the thesis choose the online tracking algorithm Deep Sort,because it is more suitable for video surveillance.3.Analyzing the reasons for the reduced tracking accuracy and insufficient robustness of the Deep Sort under the circumstances of interactive avoidance between pedestrians,and obstacles obstruction.Found that it is because of Deep Sort using the Kalman filter algorithm to predict pedestrian motion,but the Kalman filter have some defects in recognizing every situation when pedestrians are moving,lead to the estimation accuracy decrease and the error become larger in the non-linear pedestrian motion scene.In response to this problem,present a predictive method which expands nonlinear scenes,when it comes to these scenes,using the prediction results of social force model to modify Kalman filter's,and integrate this predictive method into the original Deep Sort to make it performance better.It has been confirmed that the improved algorithm does have a better performance in tracking effects through research and experiments.4.The social force model is used to simulate pedestrian motion,and it is a pedestrian motion model according to orderliness when pedestrians move.After the research of the model,the thesis applied it to the field of pedestrian tracking,to make the prediction results of pedestrian location more precise when following them.Based on the research of related algorithms,this thesis completes the establishment and realization of the model through open source data sets,and uses algorithm models such as frame Interpolation,target detection,target tracking,etc.to complete the pedestrian intelligent tracking function.