Visual SLAM System Design Based On Semantic Map

With the rapid development of Artificial Intelligence(AI)in recent years,Simultaneous Localization and Mapping(SLAM)technology,which mainly solves robot perception problems,has been widely used in many fields such as Automatic Transmission,Robot Navigation,Augment Reality(AR)and other important scenes.However,most of the existing SLAM systems are based on the geometric information,such as corners and lines.They are called low-level features compared to semantic features which are called high-level features contained in the environment.The high-level semantic information helps machine to understand the surroundings and perform better in the multivariate world,so it is very important for SLAM systems to perceive the content information in the environment from two aspects of geometry and semantics.This thesis designs a semantic SLAM system for indoor locations based on computer vision.In this system,the semantic information of surrounding environment is obtained by object detection firstly,and then Visual Odometry module will take the semantic information as input to compute the camera pose.Furthermore,the semantic information is utilized in other modules for to optimize the semantic SLAM system.Rich experiments have been carried out to analyze each module of the system,and the feasibility of the proposed system in real environment and public dataset has been verified in this thesis.Specific works are as follows:(1)SLAM front-end,also called Visual Odometry.This thesis proposes a camera pose calculation based on semantic information in the condition of the characteristics of indoor locations.In this system,different calculation method is applied to different kind of indoor locations.In the location with rich objects to be recognized,camera poses are calculated by the semantic information.In the location with fewer recognizable objects,semantic information is utilized in speeding up the feature point matching to obtain the camera pose.The experiments in this thesis prove that the pose calculation based on semantic information directly can realize the function of local mapping.(2)Loop Closing Part.According to the characteristics of the system,this thesis designs a loop closing method based on the detected objects and their relative positions.This thesis verifies the effectiveness of the loop closing method in a real environment.(3)Semantic SLAM system.This thesis builds a complete semantic SLAM system and quantitatively evaluates it on the TUM public datasets.The average trajectory errors of the two public datasets are 0.015m and 0.019m respectively while the length of trajectory of these two datasets are 7.112m and 9.263m,which meets the indoor positioning needs.