Research On 3D Environment Cognition Of Robot Based On Binocular Vision

Computer vision is an important part in the field of robot engineering.Binocular vision has developed into one of the most important research contents of computer vision.Robot 3D environment cognition technology based on binocular vision has broad development space in unmanned driving,robot navigation,industrial production and so on.Facing the binocular vision robot experimental platform,this thesis reconstructs the three-dimensional point cloud of the target object in the environment through the binocular stereo vision system,studies and designs the three-dimensional point cloud recognition network model based on deep learning,and recognizes the three-dimensional environment of the actual target object in the environment through the network.The main contents of this thesis are summarized as follows:Firstly,considering the poor matching effect and poor anti-interference ability of the traditional stereo matching algorithm under the condition of illumination distortion,a multi cost cross aggregation semi global stereo matching algorithm is proposed.The gradient information cost is added in the cost calculation process,combined with the improved Gaussian weighted mean census transformation cost as the final total cost,which reduces the influence of parallax outliers,enhances the robustness of the algorithm to illumination distortion,improves the anti-interference ability,and adds guided filtering in parallax thinning to optimize the final parallax results.Secondly,aiming at the problem that the characteristics of point cloud data and local feature extraction are not obvious,a point cloud recognition network model based on global multi-dimensional feature fusion based on deep learning is studied and designed.The network can directly process the input point cloud data,then maximize the pool of features extracted at different levels,and then map them to one-dimensional column vector features for feature information fusion,Compared with other methods,it not only reduces the amount of data calculation,but also reduces the time complexity of calculation.In the standard dataset experiment,the training accuracy is 92.8% and the test accuracy is 90.8%.Compared with other algorithms,it can achieve better recognition and classification effect,and has certain application value in the field of point cloud recognition.Then,after studying the imaging model and calibration model of binocular camera,a binocular vision robot experimental platform is built by using camera,lens,support and mobile chassis.Aiming at the specific problems that may be encountered in practical application,an adjustable spacing binocular camera and automatic dust removal system are designed,which can realize the functions of distance adjustment and dust removal.Finally,in the real scene,the target image is collected through the binocular vision robot experimental platform,and the designed three-dimensional point cloud recognition network model is tested by using the obtained point cloud data to complete the three-dimensional environment cognition.The experimental results show that the network can better identify and classify the obtained point cloud data,and can achieve the expected effect.