Research On Target Matching And Tracking Algorithms

Target matching and tracking algorithm is a hot topic of digital image processing andmachine vision research field, is widely used in the field of intelligent transportation, securitysystems, video surveillance, etc., and in other areas of our lives associated with high researchvalue and a wide range of applications.In this thesis target matching and tracking algorithmsare in-depth stdudied,and achieved optimization based on the original matching and trackingalgorithm.The thesis in target mathing algorithms studied are as follows: Firstly, studied theclassical SIFT (Scale Invariant Feature Transform) and SURF (speeded up robust features)algorithm,comparied the two algorithms and concluded that SURF works as stable as SIFT,while much faster, having better performance in real-time systems.In roder to improve the rateof correct mathcing,using SC-RANSAC(Spatial Consistency Random Sample Consensus)algorithm to reject error matching points. SC-RANSAC is faster and more robust thanRANSAC. The thesis utilize a hybrid algorithm combined SURF and SC-RANSAC together.Experiments showed that combined SURF an SC-RANSAC algorithm runs faster thanalgorithm combining SURF or SIFT with RANSAC. Secondly, studied classical LBP(LocalBinary Patterns) algorithm and PCA (Principal Component Analysis)algorithm, improved andoptimized the two algorithms, proposed face recognition algorithm based onMB-LBP(Multi-block Local Binary Patterns) operator and Multilinear PCA(MultilinearPrincipal Component Analysis) algorithm. Utilize the MB-LBP algorithm to extract facialimages, then utilize the Multilinear PCA algorithm to reduce the dimensionality of theextracted features, the nearest neighbor classifier for face recognition. An experiment isconducted, and the hybrid algorithm is proved on the FERET face database. Experimentresults show that the recognition rate of the algorithm utilized in this paper is higher than thetraditional PCA, muti-block PCA or combining them together.On dynamic target tracking algorithm mainly studied and optimized the original MeanShift and Camshift algorithm.In the aspect of Mean Shift, proposed a improved Mean Shiftalgorithm which based on the space color features and new similarity measure. First, the useof improved Mean Shift algorithm to calculate the exact location of the target in the currentframe, and then use a Kalman filter to predict an initial search position for the next frameMean Shift iteration, and finally achieved the target tracking. In the aspect of Camshift, firstimproved the classical Camshift algorithm, then fusion the Kalman filter algorithmExperiment shows that can overcome the targets color similar to the background, the targetsbecome smaller and the shield caused tracking failure.