Research On Loopback Detection Of Vision SLAM Based On Deep Learning

Visual SLAM is one of the key technologies for robot mobility and navigation,visual SLAM has a very broad application prospect.But when the camera Angle,lighting,climate,landscape and other conditions change dramatically,or when there are complex scenes with fast-moving objects,visual SLAM also has low accuracy,robustness and real-time performance,closed-loop detection is an important step to solve pose shift of SLAM,this paper presents a closed-loop detection method based on neural network.Aim at the application characteristics of closed-loop detection.Ameliorated loss function for Darknet53 network,finally,to enhance the real-time level of the system.This method is combined with binarization-autoencoder dimension reduction method,The scientific effectiveness and technical superiority of this method have been proved by experiments.The main work of this paper includes the following two aspects:(1)Aiming at the feature descriptor extracted by the closed-loop detection method based on deep learning in existence has the problems of insufficient differentiation and excessive dimension.The existing deep learning methods are analyzed in depth,Darknet53 network,which has excellent performance in time and classification,is used for substituting the feature extraction in the traditional closed-loop detection method,and starting from the characteristics of the loopback detection problem itself,improved the loss function of Darknet53 network,to better characterize the triple distance constraint relationship in closed-loop detection.Retrain the Darknet53 network with improved loss function.Finally,experiments are carried out on two open closed-loop detection data sets,New College data set and City Centre data set with more obvious changes in illumination and Angle.Experimental results show that the proposed method can detect the closed loop well and in complex scenarios,the method proposed in this paper can maintain a high recall rate while maintaining a good accuracy,and the real-time performance is also improved to a certain extent.(2)Based on the closed-loop detection method proposed above,this paper further solves the problem that the current method is not sufficiently real-time in complex scenes.Several commonly used methods of dimensionality reduction are carefully analyzed.Through the systematic study of the auto-encoder,In this paper,the thousand-dimensional feature vectors obtained by the above improved Darknet network were firstly binarized and then dimensionalized by the pre-trained autoencoder network,which greatly reduced the amount of data to be processed,thus,the real-time requirement of the whole closed-loop detection method is satisfied.Through experimental verification,the descending dimension method in this paper can achieve better dimensionality reduction effect.Finally,the Darknet-Autoencoders agile closed-loop detection method proposed in this paper is compared with other six closed-loop detection methods.Experimental results expresses that the method of this paper can achieve better comprehensive performance,higher average accuracy and better real-time performance in complex scenes.