A 3D scenes reconstruction method based on featureless block matching with a binocular vision system

• Main Authors: Chia-hua Chang, 張家華
• Other Authors: Chin-shyurng Fahn
• Format: Others
• Language: en_US
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/7gn8ru

碩士 === 國立臺灣科技大學 === 資訊工程系 === 94 === With the high development of robots, they begin to have basic thinking and recognition ability, even to imitate some works that human can do. To move accurately and safely in an unknown environment is one of the important functionality of the robots. The main purpose of this thesis is to let these robots can precisely and immediately know the locations of the objects surround them. Being different from the laser and radar sensors that are limited by the searching angle and observation distance, we use the CCD cameras to capture images to increase the robustness of our 3D scenes reconstruction system. The basic idea of this thesis is to utilize the imaging parallax of the two cameras caused by the different setup locations to accomplish 3D scenes reconstruction. While implementing this idea, we first do camera calibration such as lens distortion restoration and setup location adjustment of the cameras. Then we divide the template image into blocks and adopt the Pearson product moment correlation coefficient to find the corresponding blocks in the reference image. If the variance of the current block is too big, we further partition the block into sub-blocks recusively. We also propose an efficient searching algorithm to reduce the execution time of block matching. When we find the parallax of the corresponding blocks, we reconstruct the real 3D location by use of similar triangle calculations. The comparison of paralleled and non-paralleled cameras is also an important topic in this thesis. We prove by experiments that paralleled cameras can provide larger reconstructable area and higher accuracy then the non-paralleled cameras. Comparing with existing popular methods, our featureless block matching method not only performs faster but also better when the luminance is low or the object features are unapparent. Besides, all methods and algorithms presented in this thesis all based on the prerequisite of low computation cost to attain the most precise 3D location estimation as to help robots move immediately. The reconstruction precision of our method can reach over 90%.