Research On Human Detection And Action Recognition Technology In Intelligent Video Surveillance

From traffic monitoring to property protection,from military to civilian use,from companies to individuals,intelligent video surveillance systems have been applied to more and more applications with their unique advantages.Whether it is a safe city,smart traffic or a modern community,these new concepts cannot be separated from smart video surveillance.The purpose of intelligent video surveillance technology is to effectively detect and real-time track the interesting objects appearing in the video,so as to identify their behavior and analyze their intentions.Because of its wide range of application scenarios,intelligent surveillance systems rely on relevant technologies that have attracted the attention and research of many scholars.The purpose of the target detection is to extract the target of interest from the background in the intelligent monitoring sequence;searching for an efficient and robust target detection algorithm is the focus of scholars’ research.Target tracking is the subsequent stage of object detection.What it needs to solve is how to effectively locate the target of interest in the new monitoring sequence.In the final target behavior analysis stage,a feature-based analysis method is generally adopted: First,a feature vector that can characterize the current behavior of the monitoring object is extracted,and then it is matched with the feature vectors of various template behaviors.This thesis analyzes the related technologies used in the above processes step by step and give the improvement solution respectively.In the moving target detection phase,this thesis proposes several improvements to the traditional mixed Gaussian background model.Firstly,this thesis proposed a new background initialization method based on equal weight to reduce the dependence of the initial background model on the first frame.Then we proposed an adaptive updating weight learning rate based on state and history information.This thesis also gives a method to delete the invalid Gaussian distribution to reduce the waste of the computing resources.Finally,a motion shadow detection algorithm based on HLBP texture feature and YUV color space is proposed to optimize the moving target extraction result.In the moving target tracking process,this thesis proposes an optimized Mean shift tracking algorithm that based on multi-feature adaptive fusion and adaptive change of tracking scale,which improves the performance of the algorithm in tracking complex scenes.At the same time,this method combines the Kalman filter’s predictive function and the iterative optimization process of Mean shift method,which effectively solves the problem of poor tracking effect when the moving average speed is too fast or the moving target is blocked by an obstacle due to the traditional mean shift algorithm.In the aspect of moving target recognition,this thesis aims at the disadvantage that the traditional DAG SVMs are prone to “accumulation of errors”,and optimizes the construction process of DAG SVMs by using the distribution characteristics among the samples.Then combines the regional shape features represented by the seven moment moments of Hu with the motion characteristics of the target,the movement characteristics of the human and PCA-HOG feature to get a better feature descriptor,which improves the recognition rate of the target behavior.