| In visual servo systems, the technique of position-based pose estimation requires local invariant features extracted from the object images taken by cameras in multi-view. However, in terms of the mainstream local feature extraction algorithms, the invariance of interest point set are all ensured by complex interest point detection process and high-dimension descriptors that they cannot meet the speed requirements in industry production. Meanwhile, the smart camera is being frequently used as the processing and control center of visual servo systems, algorithms which have always been implemented on PC require to be redesigned on the new platform due to hardware and software difference.To solve these issues, this essay first propose a fast feature point extraction algorithm based on industrial scene, that foreground historgram equlization based image preprocessing, reformed SURF feature point searching and localization are introduced to improve the speed with maintained accuracy. A new local feature descriptor which can resist illumination and affine transformation is then constructed. It is composed of only 8-dimentional local moment invariant, thus much computation time is reduced in calculation of the descriptor and matching of the interest point set. To make high matching rate on this low dimensional descriptor, we designed a uniqueness constraint based rough matching and an affine constraint based Ransac algorithm during the matching procedure. At last, we conducted various levels of optimization strategies to transplant the algorithm to the smart camera platform including modified software coding, compiler configuration, and hardware acceleration by fully utilizing all the resource and characteristics of the platform.In the experiment, we testify that the algorithm can perform good invariance for illumination change and affine transformation, and reach the speed of 7 frames per second for 640×480 gray images after the final implementation on the smart camera system. |
PDF Full Text Request