Research On Object Detection And Multi-object Tracking Technology In Intelligent Video Surveillance System

With the development of computer vision technology,especially the development of deep learning technology,intelligent video surveillance technology is becoming more and more mature.As an important part of intelligent video monitoring system,object detection and object tracking technology has become an important research direction in the field of vision.This paper focuses on the research of object detection and multi-object tracking technology based on deep learning in the practical application of intelligent video surveillance system,the background,significance and research status of object detection and multi-object tracking technology are analyzed,the basic theories and methods of object detection and multi-object tracking technology are deeply studied,finally,the related algorithms are improved.In terms of object detection,according to the requirements of real-time,high precision and small memory space occupied in the intelligent video surveillance system,this paper improves the YOLO series of YOLOv3-tiny algorithm.Taking pedestrian detection as an example,by adjusting the number of horizontal and vertical directions of grid cell in YOLOv3-tiny algorithm,optimizing the network structure of YOLOv3-tiny algorithm,clustering to determine the number and size of anchor,obtaining improved algorithm,and expanding the training dataset by data enhancement method.The experimental results show that the improved algorithm significantly improves the precision of the algorithm,it is real-time,accurate and does not require much memory space,which meets the requirements of practical application of most intelligent video monitoring system.In terms of multi-object tracking,Kalman filtering can not accurate predict the objects with constantly changing motion characteristics or camera movement in monitoring system,at the same time,Kalman filtering only contains the object motion matching degree information,but not the object surface feature information.Taking pedestrian multi-object tracking as an example,this paper improves a multi-object tracking method based on Kalman filtering.The fusion model of Mahalanobis distance,cosine distance and spatial distance is used as the basis of object association.The experimental results show that the improved multi-object tracking algorithm effectively improves the SORT algorithm based on Kalman filtering,and meets the real-time and accuracy requirements in intelligent video surveillance scenarios.