Research On Motion Trajectory Cluster And Trajectory Recognition Based On Gaussian Process Regression Model

inordertoguaranteethesecurityoflivesandassets, andforthemonitoringabnor-mal situations, security monitoring systems are spreading widely nowadays. Howeverwith the drastic growing of volume of monitoring video data and information, artif-cial video analysis could not fulfl the demand of security monotoring. Meanwhile therequirement of automatical video analysis are becoming much more urgent and muchmorediversifed. Intherecentyears, theresearchofvideoanalysisbasedonvideocon-tents springs up continuously and plenty of research achievements in areas like videofeatures extraction, recognition and analysis annotation are arising, especially analysisof video contents in the security monitoring system.This thesis aims at fnding all the videos which may contain diferent part of aspecifc trajectory in a monitoring video database. The meaning of doing this is thatdiferent videos containing one trajectory may provide us various of viewpoints of onespecifc moving object which certainly will help to analysis the behavior of the object.In order to solve this problem, the thesis provides a architecture of video analysisalgorithm including reconstruction of trajectory in three dimensions, classifcation oftrajectories and recognition of classifed trajectories.In the reconstruction of trajectory in three dimensions, according to computergeometry, projection between video planes and3D space using the position of knowncorresponding points. Then the trajectory of the interested object could be reconstruct-ed in3D and similar trajectories of the object could be clustered.In the simulation process of the projection, videos with various resolution ratiosare considered and the efect of resolution of videos to the3D trajectory reconstruction are evaluated.In the motion trajectories clustering, trajectories are denoised because of the en-vironment noise in surveillance video and error in trajectories extraction. Then thetrajectories with diferent length and diferent time span are clustered using longestcontinuous point sequence distance method and modifed k-means cluster algorithm.In the trajectory recognition, using Gaussian Process Regression model to distin-guish trajectory’s category. Modeling for each point of trajectories and analysis theirsimilarity to each trajectory category according to simulation results. The robustnessand reliability of trajectory recognition algorithm using GPR model is verifed by ex-periments.