Unattended-package Detection And Behavior Understanding In Mass Transit Intelligent Video Surveillance System

In recent years,with the in-depth study of behavior understanding, unattended-package recognition has aroused extensive concern by the researchers in the field of computer vision,in particular,affected by the terrorism event,it has become an urgent need for a variety of intelligent video surveillance in public,and it is applied more and more in subway system.This thesis is based on the Shanghai Municipal Science and Technology Commission-funded project "The key electronic equipment research of rail transit network based on embedded technology"(NO. 06DZ15003).Based on the requirements analysis of the intelligent video surveillance of the rail transportation environment,we begin the study of unattended packages monitoring and recognition using image processing and pattern recognition technology.In this thesis,a method of unattended packages detection and recognition based on video is proposed.RGB subtraction,Kalman filter,template matching and HMMs are discussed,and the solutions for problems encountered in the implementation are detailed.With regard to moving target detection and segmentation,through the application research and improvement of moving target detection algorithm,and by integration with the needs of real application environment,an improved RGB subtraction method which is used to segment moving target and get the relevant information was proposed to detect moving target,To solve the problem that weather and shadows usually affect the moving target detecting,we studied the shadow detection method based on the HSV color space model.With regard to moving target tracking,in order to solve of the track of the complex environment,this paper presents a method of Kalman filter.A moving human status prediction model based on Kalman filter is built,and human outline are determined according to the result of Kalman filter tracking,In unattended-package recognition phase,First we analyze the scene to provide prior knowledge for action recognition and action understanding,and to reduce the difficulty of recognition.This was followed by the establishment of action model,the combination of target detection and segmentation to extract information from the classification of moving targets,and then define the characteristics of the various acts and pre-defined rules of unattended-package,and combined with improved methods of the Hidden Markov Model to realize the recognition of unattended-package,and optimize the state transition probability and output probability by improving on HMM training algorithm,and then improve the efficiency of unattended packages recognition.Finally,a prototype system of unattended-package detection and recognition is implemented based on VC++ and OpenCV.