Research On Image Depth Information Dataset Based On Binocular Vision

Binocular vision is an important part of computer vision.In recent years,research on binocular stereo vision matching algorithms has emerged endlessly.Binocular image datasets,as a standard for detecting and evaluating the performance of stereo matching algorithms,have important research significance.We focus on the shortage of datasets in this field and the limited number of sparse datasets for classification.We have conducted research on binocular vision principle,image dataset architecture,and sampled features to provide new test standard for evaluating binocular stereo matching algorithms.Based on the binocular vision acquisition system,we implemented a set of binocular vision image depth information dataset and evaluated the feasibility of the dataset by using three commonly-used stereo matching algorithms.The main work includes the following three aspects:(1)A binocular visual image depth information data acquisition system was set up.The system consisted of an acquisition module,a processing module,and a display module for binocular vision image acquisition and storage.The system used a laser range finder to simultaneously acquire depth information of the sample image.In this article,a double-targeted experiment was conducted.This experiment adopted a 7×9 specification chessboard calibration board,acquired 20 sets of calibration images through the image acquisition system,and successfully obtained a binocular camera calibration file based on the MATLAB toolbox and Zhang Zhengyou calibration method.(2)A set of binocular image depth information dataset was created.The dataset was divided into an image set and a label set.The image set consisted of a test set and a training set,containing 978 sets of bmp-format image data under 6 kinds of constraints and 13 types of conditions.The annotation set was a collection of XML documents.The document was used to store image information such as file names,picture names,picture sizes,laser point coordinates,and spot radius.Image sets and annotation sets establish a single association with a unique ID.(3)Assess the feasibility and practicality of this dataset.we have developed a binocular ranging system.The system integrates BG algorithm,SGBM algorithm and AD-Census algorithm to obtain matching time and depth maps.By comparing the depth information obtained by the algorithm matching with the actual depth of field obtained by the laser ranging,the matching error rate of the algorithm was evaluated.The experimental results show that the BG algorithm has the best real-time performance and the worst matching effect.The AD-Census algorithm has the best matching effect,but it takes the longest time,while the SGBM took the advantages of both the time and the matching effect.The experimental results of the matching algorithm are consistent with the theoretical research of the algorithm,which verifies the feasibility of the binocular image dataset created.