Object Detection And Accurate Localization In Industrial Application

Computer Vision is a discipline which arose in the seventies of last century. In industrial field, because the visual sensor has rich information, low cost and is simple to use, visual-based industrial task has been more important.Vision-based object detection requires us extract invariant element from the image, which is usually achieved by stable feature points or the efficient coding of image template. Object localization has closed relationship to detection, but they are different. Though we can obtain a coarse pose due to detection, but industrial applications often have strict requirements for precision. In this way, localization technology is equivalent to image registration algorithm. According to the appearance of noise, illumination changing and other factors, their exist several registration algorithms.This paper surveyed prevalent object detection and image registration algorithm. Modified the existing algorithm, designed and implemented a fast, robust algorithm for industrial object detection and a precise image registration algorithm. The main contents and results are as follows:1. Proposed a fast and robust template-based algorithm.Summarized prevalent object detection and recognition algorithms, include feature-based and template-based algorithm. In regard to texture-less situation, this paper proposed a fast method which can detect multi-object simultaneously. By training robust templates, we combine it to Hough voting framework, and this enable us to detect multi-object in 100 ms.2. Proposed a fast and accurate image registration algorithm.Summarized Lucas-Kanade and ICP framework, introduced different distance measure and optimization function. In regard to speed requirement, this paper adds additional dimension to ICP's point, using KD-tree to accelerate the algorithm, achieved sub-pixel accuracy.3. Proposed a new mathematical model which based on Graph-Matching algorithm.Summarized GA algorithm and other optimization-based framework. Modified SMAC method by using projection matrix, also implemented Graph-Matching based object detection system.