Research Of Visual Tracking Algorithm Based On Compressive Sensing

The research of visual tracking is an important part of artificial intelligence and computer vision research,which has been widely applied in fields such as video monitoring system.It also has a significant research value in practical engineering and real life.This thesis studies and summarizes the current major visual tracking algorithms,and further studies the visual tracking algorithm based on compressive sensing.Moreover,corresponding improved algorithms are proposed in order to solve the problem that tracking algorithm based on compresseive sensing is falied to track target stably for illumination changes and partial occlusion.Firstly,this thesis is detailed to introduce the compressive sensing and achieves the experimental simulation about signal reconstruction.Then the theoretical framework of visual tracking algorithm based on compressive sensing is introduced and improved algorithm based on weighted feature classification is proposed which fully take he influence of different features on classification into consideration.Through comparison experiment with original algorithm,the validity of improved algorithm based on weighted feature classification is verified.Furthermore,this thesis shows tracking algorithm based on compresseive sensing is falied to track target stably for illumination changes and partial occlusion through two experiments.When outside illumination has changed,extracting Haar feature with less calculation and Hog feature with weak sensitivity to illumination changes at the same time in the feature extraction stage is proposed.Through comparison experiments with only extracting Haar feature and only extracting Hog feature,the real-time and accuracy of improved algorithm in the process of target tracking is verified.When target has been partially covered,the idea of breaking it into smaller chunks is proposed.Making a judgement about whether the chunks are covered by Bhattacharyya distance and SIFT feature matching rate,and making rules for different chunks according to their real occulation.Through comparison experiment with initial algorithm,the validity of improved algorithm based on breaking target into smaller chunks is verified.