Localization Optimization In Visual SLAM Based On Deep Learning

Visual Simultaneous Localization and Mapping(SLAM)refers to the determination of the robot's position in the environment by the image acquired by the camera sensor and the Mapping of the environment at the same time.Traditional visual SLAM techniques mainly include feature point method and direct method,and a relatively mature technical framework has been developed.However,traditional methods rely heavily on image features.When there are motion blur,illumination change,dynamic objects in the actual situation,the algorithm is very easy to fail.With the continuous development of deep learning,unprecedented progress has been made in various fields of image processing,bringing a new approach to solve the bottleneck of traditional visual SLAM.Therefore,this paper mainly studies the deep learning method to solve the problems existing in traditional visual SLAM,so as to realize the localization optimization of visual SLAM,which mainly includes the following three contents:(1)A visual SLAM framework based on Generative Adversarial Networks(GAN)is proposed to realize the localization optimization of visual SLAM in blurring scenarios.In view of the problem that motion blurring seriously affects the quantity and quality of feature point extraction,the GAN network in the deblurring domain is introduced into the visual SLAM framework.At the same time,aiming at the defect of lack of blur detection in the existing deblurring network,the deblurring network is improved to have the blur detection function.In the process of system operation,the RGB images are firstly fed into the improved deblurring network.The improved deblurring network uses Laplacian blur detection algorithm to detect images,and deblur the blurring images to improve the image quality,and then they are fed into the localization process.This work is tested on open datasets,and the results verify the effectiveness of the new visual SLAM framework in blurring scenarios.(2)A visual SLAM system for dynamic scene processing based on segmented network is proposed.Among the existing algorithms,the traditional methods deal with the dynamic range is small,while the methods combining semantics often adopt the way of direct segmentation and culling,which can not deal with the semi-dynamic objects well.In order to solve these problems,a method of discriminating semi-dynamic objects based on semantic segmentation was proposed,which was judged by means of reprojection and intersection ratio calculation,so as to obtain all static feature points of the image and improve the robustness of the system to dynamic scenes.The results show that our method can better deal with dynamic scenes and positioning has improved.(3)Two algorithms for improving Oriented Features from Accelerated Segment Test and Rotated Binary Robust Independent Elementary Features(ORB)feature based on depth feature points are proposed.ORB,as a traditional feature point,is easily affected by illumination and Angle of view changes,and is prone to stacking.Depth feature points are often better than traditional feature points in robustness,but they are easily affected by rotation motion in matching characteristics,leading to the problem of reduction of matching point pairs.Combining the advantages of the two feature points,two improved feature points are designed in this paper.One is to screen the existing ORB feature points based on depth feature points to select high quality feature points.The other is the fusion method,which uses the depth feature points for corner detection,and calculates the corresponding corner directions and descriptors by using ORB method to obtain the final fusion feature points.The feature points are applied to the visual SLAM framework to verify the effectiveness of the proposed algorithm.