Study On Weak Supervised Semantic Segmentation In Indoor Dynamic Scene Vision SLAM System

Visual Simultaneous Localization and Mapping(VSLAM)is an important technology in the field of mobile robots which makes mobile robots more intelligent.Most of the current algorithms for visual SLAM are implemented based on static environment.If dynamic objects appear in the scene,the system of visual SLAM will be unstable.Aiming at the problem of unstable visual SLAM system in indoor dynamic scenes,this thesis combined weakly supervised semantic segmentation and visual SLAM system to remove the dynamic factors in SLAM system,which improved the accuracy and robustness of the visual SLAM system,and generated a static octotree map based on the semantic segmentation results.This article adopts ORB-SLAM2 as the basic framework of the visual SLAM system.Visual features are extracted from the input image,and the LK optical flow was used to identify the dynamic feature points.At the same time,the input image is passed into the semantic segmentation module to segment the region of the dynamic target.After removing dynamic feature points by combining the results of LK optical flow detection with semantic segmentation,the system then uses stable ORB feature points to estimate the pose of the mobile robot and to obtain a stable visual SLAM system.In the semantic segmentation of the visual SLAM system,this thesis uses weakly supervised semantic segmentation method,which greatly reduces the cost of training data annotation.And for the problem of sparse localization seeds in weakly supervised semantic segmentation,this paper applies a Saliency Guided Self-attention Network,which could guide the neural network to generate high-quality localization seeds.In order to further improve the quality of the seed,this paper also uses the adaptive color equalization algorithm to preprocess the input image,which enhances the contrast of the image,and improves the accuracy of weakly supervised semantic segmentation.In the experimental part,the visual SLAM system in this paper is verified on the public data sets TUM and in the real laboratory environment.The experiment results are compared with the classical ORB-SLAM2 system and DS-SLAM system to demonstrate the effectiveness of the method in this paper.