Research On Local Descriptor Applied To Target Detection In Video Sequence

Image local feature detection and description have received a lot of attention inrecent years. The basic idea is to first detect interested regions that are invariance to aclass of transformations. Then, for each detected region, an invariant descriptor is built.Once we have the descriptors computed, we can match interested regions betweenimages. This approach has many advantages. For example, local features can be madevery tolerant to illumination changes, perspective distortions, image blur, image zoom,and so on. Local features have performed very well in many computer visionapplications, for example, target detection, image retrieval, object recognition, texturerecognition and robot positioning. The local features of an image have a certain degreeof stability, high repeatability as well as invariance under various environments, so ithas good robustness when partial blocking, overlap, Affine transformation.This article first introduces the local description technology research status andthen introduce some related knowledge of mainstream descriptor about interested pointsextracting, descriptor generating, matching of characteristic vector set. Then, wepropose two improvements of SURF algorithm based on the theory introduced, andcheck the effect by experiments. Finally, we apply local descriptor operator to the targetdetection in video sequences, and get a better effect.Here are the main contents in this thesis:1. After analyzing the current mainstream local descriptor operators in-depth, wefind that the present algorithms of interest point extraction generally transform theimage into gray-scale image first, which leads to lose color information of the image.This paper presents the SURF algorithm based on color invariance. Experiments showthat this improvement can effectively retain the color information of the image and alsoimprove the stability of the algorithm.2. For the problem of losing image information caused by single values filled intemplate during simplifying Gaussian second derivative in the detecting link of SURF,this paper attempts to improve the values filled in the simplified template, which makesthe changing trends of simplified template closer to the changing trend of Gaussian second derivative function. The experiment shows that the improved algorithm of thispaper helps to retain more image information and extract more feature points, which canalso get more proper matching feature points between two images at the same time.3. Aiming to detect known face in the video sequence, we design a set of methodusing local descriptor in this paper. The experiment shows that the method of this papercan reduce the influence to accuracy caused by the changing expression to a certainextent, which can also decrease the computational complexity and feature matchingtime significantly. Finally, in this paper we design and implement a system of targetface detecting in video sequence.