Research On Key Technologies Of Intelligent Video Surveillance For Moving Objects

Video surveillance systems(VSS)had been widely used with the rapid growth of security needs,but the traditional VSS and technology is difficult to guarantee the promptness and effectiveness for large scale monitoring,therefore,the VSS is moving in the direction of intelligent and distributed.In this dissertation,the key technologies of intelligent video surveillance such as moving object detection,object tracking,object classification and identification are explored and studied,the accuracy and stability of the algorithm are improved as far as possible based on the existing study results,a distributed intelligent video surveillance system based on Linux was designed and implemented with Zynq-7000 SoC platform.The main work and innovation in the thesis are summarized as fallows:Firstly,Aiming at the "ghost" phenomenon generated by the target burst motion in the traditional mixed gaussian background model,an adaptive learning rate of background updating method based on three frames difference is proposed,the three-frame difference method is used to divide the monitoring image into background regions,motion region and background revealing region,the background model is updated with different learning rates for different regions.It is proved by the experiment that this method can obtain good moving target detection effect in complex motion scenes.In addition,the insensitivity of the three-frame difference method to light avoids large-scale false detection in a certain period of time after illumination mutation.Secondly,an algorithm of target tracking based on SIFT feature matching is proposed because of the rotation invariance and scaling invariance of SIFT feature.In the algorithm,a target template library is established which is used to match the detected target,and an updating and elimination mechanism based on retention priority is designed for the template library to improve the matching real-time and accuracy.In view of SIFT feature extraction algorithm is slow,we design and implement hardware accelerated logic circuit of SIFT feature extraction algorithm.It adopts the parallel mechanism inside each module,and realizes the pipeline structure among the modules,which greatly improves the speed of SIFT feature extractionThirdly,a new target classification algorithm based on SVM is proposed for linearly indivisible problems of different target features.Based on analysis of the characteristics of target classification,the SVM is trained by the shape feature:duty ratio,aspect ratio,compactness and eccentricity,which can achieve accurate classification under multi-objective condition.Last but not the least,a distributed intelligent video surveillance system based on Zynq-7000 SoC is designed and implemented to solve the problem of serious network congestion and performance degradation when the monitoring point reaches a large scale.Based on the OpenCV library,the distributed intelligent video surveillance system realizes the moving object detection,tracking and recognizing program,and judges whether the detected position and track of the different moving object break into the restricted area and warn the intruding target.