Pedestrian Detection And Tracking Of Moving Targets Based On Cascaded Visual Models And YOLO Network

Target detection and tracking technology is two very important research directions in the field of computer vision and is widely used in daily life.As the most important target,human is also the most special target.It not only has the generality of targets,but also has the variability that ordinary target do not have.Therefore,pedestrian detection and tracking technology is more difficult and challenging.This topic has both scientif c research value and social application value.In recent years,the rise of deep learning has led to many breakthroughs in the field of computer vision,especially in the accuracy of target detection.However,due to the high complexity of the depth model,the detection speed is not improved,and there are higher hardware requirements,which cannot be applied to pedestrian detection and tracking.Therefore,the first study of this thesis is pedestrian detection and tracking based on cascaded visual models,then is pedestrian detection and tracking based on YOLO(You only look once)network.Finally,the relationship between peak response and target lost is studied.So this thesis completed the following tasks:This thesis first proposes a pedestrian detection and tracking method based on DPM(Deformable Parts Model)and KCF(kernelized correlation filters).It solves the problem that the amount of calculation in the target detection is large and the real-time performance is poor,and the target tracking needs to manually calibrate the initial tracking frame.The experimental results show that the average detection frame rate of this pedestrian detection and tracking algorithm can reach more than 40 FPS(Frames per second).Based on the above research,adding the method of deep learning,this thesis proposes a pedestrian detection and tracking method based on YOLO.Firstly,this thesis trains a model for pedestrian detection.By generating the target area and correcting the target area,the bounding box of the pedestrian target can be obtained.Finally,the fusion of KCF and YOLO enables pedestrian detection and tracking.Experiments show that the method can achieve the detection efficiency of 150 FPS.It also solves the problem that about 10%of the frames are missing in using YOLO.Finally,this thesis proposes an improved KCF algorithm for determining target lost using peak response.KCF uses cyclic sampling to update the tracking template,causing KCF can not determine whether the target is lost.In this research,the KCF response of each frame in the video is extracted,and the target lost is determined by the fluctuation of the peak response.After the target is lost,a new sampling area is created by using the tracking frame information of the previous frame.A new tracking frame can be obtained by using YOLO for fast detection in this sampling area,and then the KCF tracking template can be updated.Experiments show that this method can effectively solve the problem that KCF can not determine the target lost.We set a threshold to quantify the result,this threshold limits the distance between the tracking bounding box and the actual bounding box.Finally,the precisions increased from 0.6 to 0.9 when the threshold is 50.