Study And Implementation On Key Algorithm Of Intelligent Surveillance Technology

Intelligent video surveillance is the process of describing the contents of interest in a video image by integrating digital image processing and pattern recognition, and conducting in-depth understanding and analysis. The analysis results can be used to control the monitoring system. Currently, intelligent video surveillance systems have been widely used in traffic management, public security monitoring, medical monitoring and other fields. In this paper, it involves in several key aspects of the innovative research, such as moving target detection and shadow elimination, target tracking, target classification and identification of human behavior. The specific contents and academic contributions of this paper include the following aspects.Firstly, the article studies the moving target detection and shadow elimination algorithms. In this paper, Gaussian mixture model is used to establish the background image of the video sequence and split out the moving target. To make the model converge faster, the algorithm is improved in the Gaussian mixture model initialization step. In this paper, a new approach of updating the background image is proposed. In the process of shadow elimination, the paper fully considers the color features and texture features of the moving object and the shadow, and proposes a multi-feature fusion of the shadow detection and elimination algorithm. It effectively removes the shadow of moving targets, and split out the foreground object with complete outline from the background.Secondly, in order to achieve accurate tracking of targets, the article proposes a particle filter algorithm based on incremental principal component analysis. The algorithm improves the particle weight calculation method which can reduce the degree of the particle degradation and dilution. The distances between LBP texture images of the candidate targets and the target feature subspace are regarded as the particle weights. In addition, in order to adapt to changes of the moving target, the target feature subspace is updated when the desired number of target images have been accumulated.Next, the article studied the problems of moving object classification and human behavior recognition. The moving target characterization and classification algorithm are key factors that affect the selection performance of the target classifier, therefore, this paper presents a gradient direction histograms and local texture features to describe the target, and uses SVMLight to train the target classifier. After separating traveling people from the moving target, an improved Hu moments feature vector is used to describe the behavior of pedestrians target. This paper also proposes a new moment feature vector similarity calculation method. The experimental result showed that the method improves the performance of human behavior recognition.