Research On 3D Reconstruction Of Terrain In Unstructured Environment For Mobile Robot Based On Binocular Vision

With the needs of production and life and the development of science and technology,the application of mobile robots has become more and more extensive,and the tasks facing them are more challenging.When outdoor mobile robots move on unknown and complex unstructured environmental terrain,ensuring the safety of their own operation is a prerequisite for completing the task.Through the rich information provided by binocular vision,the three-dimensional reconstruction of the unstructured environmental terrain around the mobile robot is an important way to help the robot perceive the environment,use the environment and make correct judgments,so as to ensure the stable,safe and flexible operation of the mobile robot.Therefore,the three-dimensional reconstruction of the unstructured environmental terrain surrounding the mobile robot is of great significance.First,this paper analyzes the common feature point method image matching in some scenes(barren sand,asphalt pavement,etc.),which has the problems of small number of extracted feature points and high mismatch rate,and proposes an image feature matching algorithm for discontinuous texture environment.Then Detect the points with large pixel gradient changes and Oriented FAST and Rotated BRIEF(ORB)feature points in the binocular image,and directly match the Oriented FAST and Rotated BRIEF(ORB)feature points through the Progressive Sample Consensus(PROSAC)Eliminate mismatches among them.Based on the matching rate of Oriented FAST and Rotated BRIEF(ORB)feature points,for images with low matching rate,the above retained points with large gradient changes are used,and the semi-dense direct method combined with nonlinear optimization realizes the tracking of features in the image.Through experimental demonstration,the algorithm has higher accuracy and stability in different environments.Aiming at the problems of the current sparse optical flow method and the feature point method in extracting and retaining feature information on unstructured environmental terrain,the amount of feature information is small,and the matching speed is slow,combined with the characteristics of small changes in the scale information of the binocular image,a semi-dense optical flow method is proposed The feature tracking algorithm of environmental terrain image.Traverse the entire image,calculate the gradient of the pixel grayscale change in the image,and select the pixel point with the change gradient greater than the threshold as the feature point.According to the principle of invariance of gray level,the trend of similar motion and the characteristics of small motion between frames,image feature tracking is realized.Experiments have proved that the algorithm can improve the image matching rate and stability,while retaining richer image information and reducing the loss rate of image feature tracking.Secondly,the number of three-dimensional point clouds of unstructured environmental terrain obtained is huge,and the traditional three-dimensional point cloud registration method cannot meet the requirements of timeliness and accuracy.Aiming at this problem,a method of key points in the three-dimensional point cloud of unstructured environmental terrain is proposed.Collect non-edge key points through the edge key point elimination algorithm to reduce the incompleteness of 3D point cloud description.The KD-tree accelerated nearest neighbor and iterative nearest point fusion algorithm is used to realize the error optimization after the 3D point cloud coarse registration.Experiments show that the algorithm improves the registration speed by an average of 83.238% and the registration accuracy by an average of 65.58%.Finally,the experiment platform is used to complete the 3D reconstruction experiment of unstructured environmental terrain.The image collected by the binocular camera is preprocessed,the parallax information is obtained through the principle of parallax ranging,the image and parallax information are combined to realize the 3D point cloud generation of unstructured environmental terrain,and the surface reconstruction algorithm is used to realize the environmental terrain reconstruction experiment.The results show that the experimental reconstruction results can effectively reflect the real terrain environment.