Research On SLAM Method Based On Deep Learning And Binocular Ranging In Dynamic Environment

In recent years,with the research and development of artificial intelligence algorithm,it is possible for robot to realize autonomous movement and path planning.Simultaneous Localization and Mapping are the premise for robot to realize autonomous movement.However,when there are moving objects in the environment,the traditional Visual SLAM scheme will greatly reduce the positioning and mapping accuracy.Therefore,how to improve the positioning and mapping accuracy and real-time performance of robot in dynamic environment becomes the research focus.This paper proposes a SLAM technology based on deep learning and binocular vision,which can effectively improve the positioning accuracy in dynamic environment,.Firstly,aiming at the problems of traditional binocular ranging method,such as camera calibration and stereo matching algorithm time-consuming,this paper proposes a binocular ranging method based on the combination of target detection algorithm based on YOLOv5 and radial basis function neural network,and establishes a distance prediction model.The experimental results show that this method can accurately detect and predict the distance of the target object,and the operation speed is fast.Compared with the traditional ranging method,the real-time performance is greatly increased,and the accuracy of the predicted distance can reach 97.3%.Secondly,the above distance detection algorithm is integrated into the visual odometer part,and the distance of the target object in the original binocular image collected by the binocular camera is predicted respectively,and whether the target object is a dynamic object is judged by the change of the distance between the two images,After feature extraction,it is necessary to suppress the feature points detected on the dynamic object,so that the interference of the dynamic object is eliminated in the feature matching stage,so that the camera pose accuracy of the visual odometer is greatly improved.Finally,the Bundle Adjustment algorithm is used to optimize the back-end.The experimental results show that compared with ORB-SLAM3,the relative pose error and absolute trajectory error of this algorithm are greatly reduced,and the performance is improved by about 80%.In addition,compared with the existing SLAM schemes in dynamic environment(Dyna SLAM and DS-SLAM),the overall performance of this algorithm is also greatly improved.SLAM experiments based on indoor dynamic environment prove the effectiveness of the proposed algorithm.