Research On Feature Extraction Method Of Visual SLAM Based On Fully Convolutional Neural Network

Autonomous driving is and has been one of the most prevailing robotics research topics in recent year.Researchers notices that self-driving vehicles demand a tremendous amount of research from a variety of topics in engineering,mathematics and science.Simultaneous Localization and Mapping(SLAM)technology is an indispensable part of it.Most visual SLAM methods build a sparse or semi-dense map using point features to enable real-time tracking.The monocular SLAM offers a practical and flexible approach in terms of hardware and economics.However,many monocular SLAM systems fail when facing such situations:The illumination changes drastically,moving fast that the image caught by the camera is blur or in an environment without sufficient salient features,such as white walls or glass.Recent breakthroughs in Deep Learning,especially the Deep Neural Network(DNN)based deep learning perception pipelines have been proven effective in a number of robot perception tasks.Therefore,it is necessary to deploy the most advanced deep learning model to the real robot system,which is necessary to supplement the deficiency of traditional SLAM,and the deep learning method will make the performance more robust and perfect.This article mainly focuses on the front-end problem of visual SLAM,replacing the traditional method with the feature extraction method based on neural network,and constructing a visual SLAM system on this basis.The main contents are:(1)This paper investigates the current relatively advanced visual SLAM systems based on traditional methods and learning methods,and analyzes the advantages and disadvantages of each system.The model and structure of convolutional neural network and its application in the front-end of visual SLAM are studied.It lays a theoretical foundation for the realization of feature extraction using convolutional neural networks.(2)Firstly,we calibrate the camera,calculate the reprojection error and the pose of the camera at the corresponding time,and the experimental results show that the calibration effect is good.The system structure is composed of five modules,including depth camera information reading,visual odometer,back-end optimization,loop detection and map building.Related parts of deep learning are introduced,and the convolutional neural network is summarized.PyTorch is selected as the system development framework.(3)This paper studies three traditional feature extraction methods such as SIFT,SURF and ORB,analyzes their advantages and disadvantages,and feature extraction and matching experiments are carried out.In this paper,a feature extraction method called F-CNN is proposed based on full convolutional neural network.The experimental results show that F-CNN is not weaker than these traditional feature extraction methods under the conditions of brightness change,Gaussian blur,rotation and scaling.(4)On the basis of the VGG-16 network,through adjusting the network structure and continuous training,a trained feature extraction neural network model is finally obtained.This method is compared with the three methods of SIFT,SURF and ORB under four conditions of different brightness,Gaussian blur,rotation and scale transformation.The comparison and analysis are carried out from the eleven dimensions of recall and precision.The results show that the method in this paper is better than these traditional methods.(5)Based on the previously trained network model,a visual SLAM system is constructed.Test on the KITTI dataset and compare with the latest ORB-SLAM3 system.The experimental results show that the method in this paper is better than ORB-SLAM3 on the sequences 00,03,04,06 and 10.Then it was tested in real scenes,and the results showed that the mapping effect was better.(6)At the end of the paper,the paper summarized the research,and look forward to the future work and the development of slam technology.Figure[67]table[5]reference[87]...