Pedestrian Detection And Multiple Object Tracking Using Convolutional Neural Network

With the gradual maturity of deep learning technology,this technology has been widely used in areas such as autonomous driving,intelligent monitoring and humancomputer interaction,including drones.In this thesis,detection and tracking some pedestrians as a practical application,and the two most important parts of this task are studied,they are pedestrian detection and online multiple object tracking.Relying on deep learning technology,thesis research on pedestrian detection and tracking algorithm based on convolutional neural network.For pedestrian detection,first of all,this thesis selects the YOLOv3 network model with high detection accuracy and good real-time performance as the pedestrian detection model through comparative analysis.Use pedestrian detection set to study the detection model,this thesis collects and organizes the pedestrian images and marks.To solve the problem of low detection accuracy,three optimization methods are used,which are pedestrian sample classification,K-means clustering to obtain the highest priority candidate frame and hard example mining,which successfully improved the pedestrian detection accuracy to 90.89%.Secondly,in view of the problem of low detection accuracy of far-field pedestrian,this thesis proposes the improved YOLOv3 model,which improves the backbone network structure of the YOLOv3 model and merges some residual unit structures in the network.Using the far view data set as a verification,compared with the YOLOv3 pedestrain detection results,the improved YOLOv3 model pedestrain detection accuracy m AP(mean Average Precision)increased by 4.53%,and the pedestrain detection precision Io U(Intersection-over-Union)increased by 7.02%.This thesis preprocess the pedestrain pictures in the classic data sets to generate auxiliary data sets,and import it to different detection model for training and testing,comparative result of the advantages and disadvantages of improved YOLOv3 model and other deep learning models.It embodies the excellent pedestrain detection accuracy and robustness of improved YOLOv3 model in various scenarios.For the pedestrain tracking process,on the basis of highly accurate detection of improved YOLOv3 for the pedestrian.This thesis aims at online multiple object tracking,Firstly,this thesis analyzes the architecture and key technologies of multiple object tracking.Secondly,according to the characteristics of motion,shape and depth appearance,a high-confidence hierarchical data association model based on multi-information fusion is proposed.Then,correlation filter is used to calculate the confidence of the associated pairs,which improves the tracking effect and processing efficiency of the tracking model.Finally thesis conduct comparison experiments on tracking public verification sets to analyze the tracking effects of different models.Compared with other multiple object tracking models,the multiple object tracking model proposed in this thesis performs better,in the performance of the model to balance the tracking accuracy and processing time.