Research On Visual SLAM Algorithm Based On Convolutional Neural Network Multiscale Feature Fusion

In recent years,with the rapid development of mobile robots and autonomous driving,synchronous localization and mapping(SLAM)technology has received widespread attention.Among them,visual sensors have gradually become the most concerned research direction in the SLAM field due to their low cost and strong adaptability.However,most of the current visual SLAMs use traditional manual design algorithms to extract image feature points,which has significant limitations for complex scenes with dynamic and varying lighting conditions.With the continuous development of deep learning technology,it has been proven that deep learning technology has strong robustness and adaptability to complex environments in different visual tasks.Through deep learning technology to extract the deep semantic information of the image,we can obtain a stronger feature representation of the image itself,which proposes a new strategy for improving the performance of the visual SLAM system.This thesis analyzes feature point extraction algorithms based on deep learning,and studies the fusion of traditional visual SLAM algorithms and deep learning feature extraction methods.The specific work completed is as follows:(1)Firstly,traditional feature extraction algorithms and deep learning based feature extraction algorithms were introduced and analyzed.Three feature extraction modules,ORB,Super Point,and D2-Net,were constructed respectively.In relevant experiments,the three feature extraction algorithms were compared to verify the effectiveness of deep learning convolutional neural network based feature extraction algorithms in feature point extraction.(2)Secondly,in order to improve the feature location and shape perception ability of the convolutional neural network,the network structure of D2-Net is improved,and a multiscale attention module and an attention guidance module are proposed to fuse the low-level structural information and high-level semantic information of the network,and the traditional convolution and deformable convolution results are weighted to guide to reduce the damage to the image structure.The convolutional neural network designed for multiscale feature fusion was evaluated in multiple tasks,verifying the effectiveness of the improved network’s localization and shape perception capabilities.(3)Finally,in order to improve the localization and mapping capabilities of the visual SLAM system under different scene conditions,the multi-scale feature fusion convolutional neural network feature extraction algorithm designed was fused with the traditional visual SLAM algorithm,and a new map construction thread was designed to add a dense point cloud construction thread to solve the limitation of the visual SLAM system only generating sparse point cloud maps.In the experimental section,a visual odometer experiment was designed complete SLAM system performance experiments and real scene experiments.By comparing the proposed algorithm with ORB-SLAM2,GCN-SLAM,and DXSLAM algorithms,it was found that the overall absolute trajectory error and relative pose error were superior to the compared algorithms.Moreover,the dense point cloud map of the constructed scene was not significantly different from the experimental scene,which verified the effectiveness of the map construction thread in real-world scenarios.