Research On Target Detecting And Tracking On Low Rank Representation

Low Rank Representation(LRR)is a method to decompose the observed data into the form of low rank block and sparse block,which is concerned by the majority of scholars because of its good robustness to noise.So far,low rank representation method has been applied frequently for face recognition,target detection,etc.and achieved good results.In addition,in recent years,some scholars attempt at application in target tracking with low rank representation method and fortunately made some achievements.Low rank representation method is somewhat similar to the "Tiger Balm".And,specific issues still need concrete analysis.The LRR often needs to combine with other theoretical knowledge,so as to obtain better algorithm for achieving the optimal performance.1?Relevant theoretical analysis.Firstly,the theory of low rank representation is analyzed and discussed,and then the related theory of particle filter used in moving target tracking algorithm is described and analyzed.2?Dynamic updating projection via low rank representation for online moving objects detection.After analyze several traditional moving detection methods,a dynamically updating projection method for online moving objects detection is proposed.The proposed method use LRR method to obtain low rank part of several continuous video images,and thus the projection can be constructed with orthogonal complement of the left singular matrix from the obtained low rank part,then sparse foreground can be obtained by solving the projection model,what's more,the video can be divided into several uniformly-spaced part,based which the projection can be dynamically updated.Experiment results on several video databases confirm that our method has better detection performance than other methods,and especially in dealing with dynamic background and complex foreground,the proposed method has strong robustness.3?Single target tracking method based on weighted low rank representation.Under framework of particle filter,we propose to apply weighted low rank representation for single target tracking.With the proposed method,the target tracking problem is transformed into low rank representation model,and the model is solved with weighted low rank representation method.And then,low rank representation coefficient corresponding to each particle on dictionary templates can be obtained.Since dictionary templates are selected by the combination form of target templates and background templates,then for each particle,difference between sum of coefficients on target templates and sum of coefficients on background templates can help make decision,the greater the value,the greater the probability that the particle as the target.What's more,as for the update of the template,we adopt similarity measure method with track results and target templates,when their similarity is greater than the similarity threshold,we update the template,or the template will not be updated.The experimental results show that the proposed method is robust for change of scale,illumination,occlusion,and posture.