Visual SLAM Research Based On Deep Learning

VSLAM is a process in which a robot uses a camera as a data acquisition sensor to perceive the unknown environment around it,so as to realize self-positioning and map construction.As one of the key technologies of mobile robots,VSLAM has become a research focus in recent years.Most VSLAM algorithms are based on geometric structure to process data information,unable to obtain higher-level information,which reduces the robot's perception and understanding capabilities.The development of deep learning provides a direction to solve the above problems.Combining deep learning with traditional VSLAM,understanding the scene from the two layers of semantic and geometric structure information can make up for the shortcomings of traditional VSLAM.This paper uses the traditional VSLAM algorithm ORB-SLAM2 as the basic framework,and optimizes the traditional SLAM based on the deep learning algorithm.The specific research content is as follows:(1)Aiming at the problem of low accuracy of pose estimation in traditional SLAM algorithm in dynamic environment,a visual odometry method based on semantic segmentation is adopted.This method firstly extracts image features through ORB algorithm,then uses deep learning-based PSPNet semantic segmentation algorithm to segment the dynamic objects in the image,then removes dynamic features,and finally calculates the pose through feature matching and ICP algorithm.Experiments prove that the method in this paper effectively eliminates dynamic features and improves the accuracy of pose estimation.(2)Aiming at the low efficiency of traditional SLAM algorithm using bag-ofwords model for loop closure detection,a fast loop closure detection method based on Mobile Netv3 is adopted.This method uses the Mobile Netv3 convolutional neural network based on deep learning to extract the features of the image,and reduces the dimensionality of the features through the PCA algorithm,and finally determines whether a closed loop is generated through the cosine similarity.The comparative experiment proves that the loop closure detection method in this paper is fast and accurate.(3)Aiming at the problem that traditional SLAM algorithms cannot construct semantic maps,a method of constructing semantic maps based on object detection algorithms is adopted.In this method,the point cloud data of key frames is segmented through the super-voxel clustering algorithm to generate a preliminary semantic map.Then,the key frames are used to obtain the semantic information of the objects in the image using the YOLOv4 object detection algorithm based on deep learning,and construct the point cloud with semantic labels,and finally merge the preliminary semantic map with the point cloud with semantic labels to build the final semantic map.Through experiments,a hierarchical and clear 3D semantic map is constructed,and the objects in the scene are marked with different colors,which improves the perception ability of the visual SLAM system.