Research On Target Tracking Algorithms Based On Compressive Sensing

Stereo matching and target tracking are hot issues in study of machine vision,and are widely used in intelligent transportation,visual navigation and medical diagnosis.But there are many problems have not been resolved,of which the main difficulties are from the calculations of algorithms,occlusions,fast target moving and so on.Based on the theory of compressive sensing,this thesis proposes a fast stereo matching algorithm and target tracking algorithms with high robustness and real-time performance by introducing multi-feature fusion and object position prediction in machine vision.The main work includes the following three aspects:(1)In order to reduce computation and obtain denser disparity map,this thesis presents a fast quasi-dense stereo matching algorithm for calibrated images.Firstly,the image feature points are extracted by the traditional SIFT algorithm.Then the dimensions of the SIFT feature vectors are reduced by the sparse random projection of compressive sensing.Finally,the matching points are spread as the seed points to the whole image,and the dense disparity map is obtained.The experiment results show that,compared with SIFT algorithm with the seed diffusion algorithm,compressive SIFT algorithm with seed diffusion algorithm can reduce the complexity of the algorithm and ensure the quality of the disparity map.(2)In order to improve the robustness and efficiency of tracking,this thesis presents a compressive tracking algorithm which combines color and texture features.Different from traditional algorithms,it extracts the target’s color and texture features.Then two projection matrices are used to map the LBP texture feature of the image and the color feature of the H space to the low-dimensional space.Finally,the background weighting method is used to fusethe two features of the compressed domain.The experiment results show that,compared with the performance of traditional compression tracking algorithm,the proposed algorithm has higher tracking accuracy and better anti-similarity and anti-occlusion interference.(3)In order to achieve good tracking effect in a complex scene for the fast moving target,this thesis presents a multi-feature fusion compressive tracking algorithm with Kalman target prediction.Firstly,Kalman filter is used to predict the position of the target in the next frame.Then the distance between the sample position of compressive tracking and the predicted position of Kalman filter is calculated.Finally,the location weights are exploited in the Bayes classifier for classification.The experiment results show that,compared with traditional compressive tracking algorithms and Kalman filter based Meanshift algorithms,the proposed algorithm has higher accuracy and faster tracking for fast moving objects in complex scenes.