Extended Object Tracking Based On Feature Extraction And Feature Description

As one of the important part in digital image processing and pattern recognition, object tracking has extensive applications such as video surveillance, video retrieval and behavior analysis. The traditional object tracking usually uses the difference of objects and backgrounds in grey information to segment the object and obtain its location information. However, for extended targets, they usually occupy a large of proportion of the view field and are influenced by some factors including pose change, scale variation, occlusion and so on in the process of motion, which cause partial features information of object missing and feature mismatched, then the tracking results will be not correct. So, it is still a challenge to design a robust tracking algorithm to achieve the goal of the robust tracking for extended targets.This paper concentrates on feature description and finds the most effective mean of object representation. Aiming at the disadvantages of existing feature description and object tracking algorithms, it proposes some effective and new methods. At first, this paper analyses the SURF and BRISK descriptors, and proposes a new feature description SURF-BRISK, getting rid of the non-real time of SURF and the quantitative limitation of feature points of BRISK. This algorithm uses the SURF descriptor to obtain the feature points, and the BRISK descriptor to calculate the descriptor, enhancing its real-time as well as keeping the capability. This paper uses the primary existing description means for target location. First it obtains the feature point. Then, the mistake match is wiped off by RANSAC algorithm. Finally, according to the correct match points, it calculates the affine matrix and finds out the object location in the next frame. It is shown that the SURF-BRISK algorithm is better than other description means.The object tracking algorithms based on the character match are unable to handle the video with few textures, so this paper proposes an algorithm based on the local sparse appearance model——LSAM. First of all, it introduces the particular theory of particle filter. Then it describes the used tracking frame of particle filter. Finally, it demonstrates the theory of the proposed algorithm in detail. The proposed algorithm uses sparse representation and local image block overlap sampling to construct the appearance model of object. By the equalization step, it gets the vector expression of the object. Combining the sparse representation with the increasing quantum space algorithm, it updates the model. In the Bayesian frame, tracking is regarded as the problem of calculating the maximum posterior probability. The frame of whole algorithm composes of four modules including dictionary construction, sparse appearance model of candidate object, maximum posterior probability calculation and model update.This paper uses a lot of classical tracking videos for algorithm test, and compares it with the existing algorithms objectively. Experimental results have shown that the proposed algorithm performs better than other tracking algorithms. However, there are also some disadvantages that real-time is not high, and tracking effect is not especially good when the object appears motion-blur. Finally, this paper summarizes and analyzes the study contents and innovative points comprehensively, figuring out the new thoughts of the algorithm in further investigation.