Research On Monocular Visual SLAM Algorithm Based On ORB Feature

SLAM is a practical robot positioning and mapping technology,and will make the robot quickly familiar with the surrounding environment and is helpful for robot to realize the autonomy and intelligence.Monocular is more and more popular in visual SLAM applied on the mobile terminal due to its advantages such as small amount of data collected,convenient installation and low price.However,one of the shortcomings of monocular cameras is that they cannot obtain the depth of pixels in the image,and most pixel depth estimation methods have the problems of large computation and long inference time;at the same time,the ORB feature point has poor adaptability to the environment with large light changes.In this paper,a monocular visual SLAM algorithm based on vision ORB features is designed by referring to the ORB-SLAM2 algorithm framework,at the same time,the effects of the estimation of the depth of monocular images and the ORB feature extraction algorithm on the performance of monocular visual SLAM are further studied.(1)This paper aims to solves the problem that current monocular camera cannot directly obtain the pixel depth.A fast monocular image depth estimation network based on multi-scale feature fusion is designed by referring to deep learning methods that are widely used today.Ghost Net is used as the coding network of the monocular image depth estimation network to improve the coding speed of the network;deconvolution and bilinear interpolation is used to design the decoding network,in the meantime,characteristics of the coding network with the characteristics of the decoding network are merged through a cross-layer connection to enhance the edge details of objects in the depth map.This network is integrated into the initialization of monocular visual SLAM.A monocular visual SLAM algorithm based on CNN depth estimation is designed and experiments are carried out.The experimental results show that the initialization speed of the monocular visual SLAM algorithm based on CNN depth estimation is faster and has an impact on the accuracy of the camera trajectory.(2)In terms of feature point extraction,ORB feature extraction algorithm has poor adaptability to environments with large changes in light intensity,the threshold in the original ORB feature extraction algorithm is calculated using a local gray-scale adaptive threshold algorithm to improve the original robustness of the algorithm.Based on the improved ORB feature extraction algorithm,a visual odometer and CNN depth estimation network were designed,and then a monocular visual SLAM algorithm based on vision ORB features was designed The experimental results show that the proposed algorithm is slightly better than ORB-SLAM2 in precision.