Multiple Running Vehicles Tracking Technology In Video Frame

The detection and tracking of Moving targets is an important component of Computer Vision, which is attracted a large number of domestic and foreign scholars, and it is also a technology foundation of realizing the Intelligent Traffic System. Because of the low efficiency of traditional tracking methods which based on model and regional, this paper is devoted to the research of the tracking algorithms which based on feature extraction. In this numerous algorithms, SIFT (Scale Invariant Feature Transform) which is proposed by David Lowe has shown the greatest advantages in feature extraction. They are invariant features, illumination invariant characteristics, scale invariance, affine invariant. First, this paper introduces the domestic and foreign research status of targets tracking especially vehicle tracking in detail, then, analyzes advantages of the SIFT algorithm with the other ones which are based on feature extraction, and introduces SIFT algorithm's foundational theories: Gauss Convolution, Scale space with some examples. Then, this paper takes the SIFT algorithm steps as the main line and combined with the actual experiments to analysis SIFT in detail. SIFT's main contents are building the Gauss gold tower of images, the reason of using DOG operator and DOG operator's characteristics, positioning of feature points and filtering operation of feature points, calculation of gradient direction and magnitude of feature points, generation of feature point's description operator vector by neighborhood pixel gradient information, the improved K-D algorithm--BBF search tree, clearing of mismatch points by RANSAC. This paper analysis all of this contents by corresponding theories and examples.A very important content of this paper is the analysis and discussion of classic improving algorithms based on SIFT:PAC-SIFT, SURF, and ASIFT. And, this paper also compares those classic improving algorithms with SIFT in performance and efficiency.This paper improves SIFT algorithm from three aspects with relevant analysis of experiments, they are:1. after build of image's DOG Pyramid, the method of determining prospective-keypoints, namely:Are that points really the closest to the true key point in Sub-pixel? And this improvement is based on a dynamic coefficient selection.2Determination of the threshold of Euclidean distance ratio and similarity of vectors when determining whether the two matching points is correctly. this paper uses an adaptive dynamic threshold algorithm to determine the threshold, and vector space cosine theorem to solve the problem of vector similarity.3Big error matching in SIFT is the always exist, and if using of the Big error matching points for the subsequent RANSAC mismatch clear work will make the algorithm becomes redundant, and precision drops, therefore this article drops Big error matching points out before the staring of RANSAC algorithm by a kind of absolute slope calculation method. And this paper also gives the related computing algorithm and analysis of some examples.Finally, this paper proposes a tracking system of moving-vehicles which based on this improved SIFT algorithm, and the experiments results show good.