Accurate Fast Localization Method And Research On Human Pose Analysis

Computer Vision is discipline which arose in the seventies of lastcentury. In industrial field, because the visual sensor has richer information,lower cost and is simple to utilize, visual-based industrial task has beenplace at important position.Vision-based object detection requires us extract invariant element fromthe image, which is achieved by stable feature points or the efficient codingof image template. Object localization has a tight relationship with detection,but they are just different. Though we can obtain a coarse pose fromdetection, but industrial applications often have a very strict requirement forprecision. In this way, object localization technology is equivalent to imageregistration algorithm. According to the affect of noise, illuminationchanging and others, there exist several registration algorithms.This paper surveyed many prevalent object detection and imageregistration algorithm. We modified up-to-date algorithms, designed andimplemented a quite fast, robust algorithm for industrial object locaolizationalgorithm. The main content and result are listed as follows:1. Proposed a novel Shape Context descriptor for object localization.Based on others’ work, for the situation of lacking texture on thesurface of object, we proposed a novel Shape Context descriptor.By introducing high-level sampling, a higher dimensional histogramrepresentation and extra angle of point, we achieve to make a precise description of points even if on textureless object.2. Proposed a fast and accurate image registration algorithm.We summarized Lucas-Kanade and prevalent ICP framework,introduced different distance measure. In consideration of speedrequirement, this paper adds additional dimension to ICP’s pointdescriptor, using KD-tree to accelerate that algorithm, and finallyachieved sub-pixel accuracy.3. Proposed a framework of human pose estimation algorithm, aimedat a monolar view.We abosorb the framework of multi-view human pose estimation,and try to adjust it to monolar view situation, where human bodystructure can be hardly got from one single picture. This frameworkcan be used for successive human body reconstruction and tracking,and should become necessary for motion and behavior recognition.