Font Size: a A A

Research On Visual SLAM Method Based On Deep Learning

Posted on:2022-10-13Degree:MasterType:Thesis
Country:ChinaCandidate:C XueFull Text:PDF
GTID:2518306347481344Subject:Master of Engineering
Abstract/Summary:PDF Full Text Request
Nowadays,intelligent robot technology is developing rapidly,and simultaneous localization and mapping(SLAM)technology is also constantly improving.With the improvement of the performance of visual sensors such as cameras,visual SLAM technology has increasingly become a hot research field.At the same time,with the application of deep learning technology in different fields and outstanding achievements in image processing,the combination of visual SLAM and deep learning to improve the performance of a certain module in the SLAM program has become a research hotspot of related researchers.This paper analyzes the current research status at home and abroad,based on the traditional visual SLAM combined with deep learning,and carried out the research on the visual SLAM method based on deep learning.The main research contents of this article are as follows:The first step is to integrate the convolutional neural network framework of the visual odometer.The depth estimation network based on Resnet50 and DispNet is combined with the Liteflownet optical flow extraction network for joint training,and the optical flow extracted by the optical flow network is passed through the opposite pole.Geometry and other methods calculate the camera motion pose,and then estimate the depth of the scene through the triangulation module,and adjust the scale consistency between this depth and the depth obtained by the depth estimation network,so as to design the corresponding loss function to train the network to make the final generated The model can be improved in accuracy and verified in experiments.Second,I then studied the combination of the loop detection module in the visual SLAM and the autoencoder method of deep learning.First,grayscale the training input image,and the image was divided into two identical copies,and one input was input to the convolutional autoencoder.The encoder performs feature extraction and reconstruction,and the other is randomly adjusted for motion blur,defocus blur,and brightness,and then input into the histogram of gradient direction(HOG)feature extractor to obtain the extracted multi-dimensional feature expression.Perform a two-norm loss function design with the features reconstructed by the autoencoder,perform network training to obtain the model,and then establish a feature storage library,and perform similarity matching on the input new image frames through simple linear search,and determine the loopback.Finally,this article designed the above two experiments in a reasonable way,trained the model,tested the trained model in the visual odometer experiment,and output the result graph of optical flow extraction,which is qualitatively compared with FlownetC,FlownetS,and Unflow Compare,and then output the depth estimation map,supplemented by the monocular depth estimation evaluation index and DF-Net,SfinLeamer,Monodepth2,Zhao and other schemes for comprehensive analysis,and the visual odometer on the KITTI odometer data set for trajectory evaluation,simultaneously test in the actual environment,from Many aspects have confirmed the accuracy and robustness of the proposed scheme.For the loop detection experiment,mainly by inputting three different data sets into the trained model,through feature extraction and retrieval,the image similarity judgment is performed,the loop frame is obtained,and the PR curve and AUC of the system on the different data sets are output.The area and detection speed improve the loop detection performance from the loop accuracy and speed.
Keywords/Search Tags:Visual odometer, Loop detection, Depth estimation, Convolutional autoencoder, HOG feature
PDF Full Text Request
Related items