Vision Based Object Detection, Analysis And Description

Inspired by booming computer science and artificial intelligence, peopleare trying to design such a vision system that the same as that of the humanbeings, which is precision, complicated and able to understand the world.Above is the major task of the computer vision subject. In recent years, theinternational academic and business circles in this area has made extensiveand in-depth study, and achieved remarkable progress and development. Thework done by this thesis not only related to the bottom of this subject, but al-so for the high-level semantic analysis. From a low level perspective, thework consists of image data capturing and saving, semantic region extractfrom images and videos. Base a high level view, this thesis is aimed to recog-nize the feature of the object, the trajectory and behavior. It has many poten-tial applications in smart surveillance, human-computer interaction, sports,health, video codec etc.The major research contents and results of this thesis are as follows.(1) In this paper, we proposed a workable solution for the camera cap-ture physical layer (hardware architecture) and data acquisition and storagelayer of the markerless motion capture in a middle level computer.300M datacan be acquired and stored stably per second.(2) We proposed a novel method for detecting salient region detection from images based on human saliency mechanism. The Experiments donebased on the public dataset shown the proposed method was effective even inthe images which having a complicated texture background. By computingthe saliency map, our model can rapidly orientate towards salient objects in acluttered visual scene, which may reduce the vast amount of incoming visualdata.(3) This paper presented a novel method for detection and recognition ofglass defects in low resolution images. First, the defect region was located bythe method of Canny edge detection, and thus the smallest connected region(rectangle) can be found. Then, the binary information of the core region canbe obtained based on a specific filter. After noises are removed, a novel Bi-nary Sequence Pattern (BSP) was proposed to describe the characteristic ofthe glass defect. Besides, another two features, symmetry and length-widthratio are also employed in the proposed algorithm. Finally, the AdaBoostmethod was adopted for classification. Experiments with800bubble imagesand240non-bubble images proved that the proposed method was effectiveand efficient for recognition of glass defects, such as bubbles and inclusions.(4) A method for multiple objects detecting and tracking was proposedbased on the extracted foreground. In the large-scale movement of objects,the shape and texture of each individual are very similar. The region andcontour information were used to track object between two frames. Shapesimilarity description method is proposed. Efficient algorithm is used to fur-ther improve the level of real-time tracking.