Research On Monocular Slam Algorithm Based On Deep Neural Network Feature Extraction

The rapid development of computer vision technology promotes the popularization of intelligent mobile robots.Simultaneous Localization And Mapping(SLAM)is a key technology for robot localization and navigation in unknown environments and has attracted extensive attention from domestic and foreign researchers.SLAM technology mainly solves the localization and mapping problems of robots equipped with sensors without prior environmental information.Laser SLAM and visual SLAM are two major research directions in the field of SLAM.Visual SLAM based on feature point method is the hotspot of current research.However,the feature points extracted by feature extraction algorithm used by this kind of algorithm are relatively single with low robustness,which will affect the accuracy of subsequent pose estimation.The method based on deep learning can extract more correctly matched feature points than the traditional method,and replace the features in SLAM with deep learning features,which will alleviate the problem of poor performance of feature points.The feature points extracted by the feature extraction algorithm used in the traditional visual SLAM algorithm are less robust to illumination changes,viewing angle changes and noise.This paper replaces the features in the traditional SLAM algorithm with deep learning features,designs a feature extraction module based on Super Point,optimizes the designed feature extraction module based on image pyramids,and use a two-way nearest neighbor matching to improve the feature matching module,propose a monocular SLAM algorithm based on deep learning features.The analysis of the experimental results shows that replacing traditional features with deep learning features can improve the accuracy of SLAM trajectory estimation and reduce the error in the pose estimation process.The improvement of feature extraction and feature matching modules can further improve the accuracy of the trajectory.The feature extraction module based on Super Point will produce dense responses to images with rich details,which will increase the mismatch rate and affect the accuracy of subsequent pose estimation.At the same time,the visual odemetry will produce cumulative errors when performing pose estimation,which affects the performance of the algorithm in large-scale scenes.For this reason,this paper designs a feature extraction module based on quadtree optimization strategy to manage the extracted feature points and obtain globally uniformly distributed feature points.In order to improve the trajectory accuracy of the algorithm in large scenarios,this paper proposes a loop closing module based on the deep learning feature bag of words model.This paper uses the image dataset Bovisa to train the visual feature dictionary.By clustering the Super Point features extracted from each image in the dataset,the offline training of the feature dictionary is completed,and the visual feature dictionary is constructed as a K-ary tree to reduce reduce the time spent in word search.According to the analysis of the ablation experiment results,the implementation of the quadtree partition optimization strategy for the feature points reduces the RMSE(ATE)index;the visual feature dictionary can be effectively used in the loop closing module to eliminate the cumulative error of the pose estimation.Compared with other existing methods,the monocular SLAM algorithm proposed in this paper has accuracy advantages.