Research On Multi-object Detection And Multi-object Tracking Algorithm In The Monitoring Video

Car ownership in China is in a state of rising year by year,auto market in the outbreak,and at the same time the problem of road traffic is becoming more and more serious.The wisdom of the Chinese urban construction,with the popularity of road monitoring equipment,as well as the improving of the surveillance video image quality,improve road surveillance video data storage,from TB level to the level of PB,how to efficient use of these data,reduce the storage costs,is an important topic.It is more useful to use the important information extracted to the city management to realize the intelligence of city management.This paper studies a variety of cutting-edge technology and designs a complete set of multi-object detection and multi-object tracking system in surveillance video.The work mainly includes two parts.First,the road monitoring video target detection and classification,that is,the target detection algorithm research and detector design,we focus on three categories,the car,people and motorcycles.The purpose is to obtain the location information,and determine its type.Then determine the categories of objectives.This part mainly uses the deep learning technique,and uses the deep regression neural network to carry on the target detection.Second,multi-target tracking algorithm.Mainly studied the KCF tracking algorithm based on kernel related filtering algorithm application in surveillance video,and design the prediction correction module,using kalman filter algorithm for tracking the real-time correction.Our system combines detector and the tracker.Using the detector designed in the first part,to obtain the location of the object the category information to initialize the tracker.For multiple target tracking,and connecting with the subsequent detector of continuous testing,data associated with the tracking data,to correct the tracker,with exception handling technology,eliminate missed detection,error detection,tracking lost etc,to ensure accuracy and robustness of tracking.In the end,the performance test of the complete system was performed in the actual monitoring scenario,and the detector and the tracker met the performance indexes of the intended design.