A hardware implementation of disparity estimation based on edge preserving

• Main Authors: Lin Ding-Xiang, 林鼎翔
• Other Authors: Shiau Yeu-Horng
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/kh7x3d

碩士 === 國立雲林科技大學 === 電機工程系 === 107 === After the advancement of mechanical technology in the industrial age, robots can operate in an environment in which humans cannot survive. Machine vision systems are an indispensable technology in achieving autonomous operations. Stereo vision is a technology that provides depth information in machine vision technology. The technology uses two cameras to simulate human eyes to capture scenes separately. Then, the two captured scenes are used to calculate the distance between the scene and the camera. Such method is not easily affected by the external environment. However, this kind of system needs high processing performance. If the system is used as a sensor, it means that this technology needs to have both accuracy and immediacy. This thesis uses adaptive support weight (ADSW) to estimate the disparity map. The proposed method uses the weighting operation to find the best displacement point, and uses the sparse statistical transformation (Sparse Census) to reduce the disturbance of light source, which can provide good performance in terms of accuracy and speed. In addition, in order to preserve the edge information, image pre-processing and post-processing are added. For pre-processing, the edge sharpening filter (edge crispening) is first adopted to the input image. That is to optimize the image to obtain a correct weight value in the discontinuous region. For post-processing, the edge detection is applied to the estimated disparity map to reduce the loss of edge information. In order to implement the hardware circuit of the algorithm, an algorithm for optimizing the hardware is established. The exponential operation is replaced by a special shift approximation method, and the accuracy of the hardware algorithm is improved by amplifying the calculated value to avoid obtaining multiple identical weighting values. We used software toolbox in OpenCV to simulate the algorithm. To achieve the iii requirement of real-time applications, we designed hardware architecture of the proposed method by using Verilog, and then verified the architecture by Quartus II. In the experiment, we estimated the error rate with Middlebury website. The software average correct rate is 92.58%, and hardware average correct rate is 90.94%.