Font Size: a A A

Research On Visual Position Recognition Technology In Large-scale Unstructured Environment

Posted on:2021-03-22Degree:MasterType:Thesis
Country:ChinaCandidate:J L HaoFull Text:PDF
GTID:2518306464477464Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
Position recognition technology is an important research area of computer vision.It is a key component of loop detection in Visual Simultaneous Localization and Mapping(VSLAM).It is widely used in robot navigation,automatic driving and other fields.In the research of position recognition technology,it is of most practical significance to realize effective recognition under the condition of large viewing angle and apparent changes.In large-scale unstructured environments,there are large changes in perspective and appearance,and traditional feature extraction methods are no longer applicable.So this paper proposes a visual position recognition method based on deep convolutional neural network.By introducing the attention mechanism,the visual position recognition network has higher accuracy and robustness in the large-scale unstructured environment.On the experimental platform equipped with GPU,the verification and comparison experiments performed on the self-built KL data set and the public data set(KITTI data set and SL data set)show that our proposed image feature extraction method is in a large-scale unstructured environment More robust.The main research results of this article are as follows:1.Aiming at the problem that traditional feature extraction methods(such as SIFT,ORB,and BRIEF,etc.)cannot be used in large-scale unstructured environments,an image feature extraction method based on deep convolutional neural network Res Net is proposed to solve the problem in large-scale environments Facing the problems of light extraction,camera shooting angle changes,feature extraction when there are moving objects,buildings,and ground surface changes.2.In view of the lack of attention link in Res Net,the attention link SE-net is introduced into Res Net to improve the robustness of feature extraction.3.This paper uses Google Street View Time Machine and Python crawler to build a million-level high-quality visual position recognition dataset,which solves the problem that no large-scale visual position recognition dataset is available.4.Aiming at the problem that the existing loop detection algorithm in SLAM fails in a large-scale unstructured environment,this paper applies visual position recognition technology based on deep convolutional neural network to ORB-SLAM for loop detection,which effectively improves ORB-SLAM.Control accuracy,robustness and operating efficiency in large-scale non-structured environments.5.Based on the above,a visual position recognition application software is written.
Keywords/Search Tags:Visual Position Recognition, ResNet, SE-Net, Google Street View, ORB-SLAM, Loop Detection
PDF Full Text Request
Related items