Research On Real-time Visual 3D Construction For Mobile Robot Localization And Navigation

Mobile robots are widely used in various scenarios such as indoor logistics,field detection,and air flight.Visual Simultaneous Localization and Mapping(VSLAM)is considered to be one of the important foundations for mobile robots to perceive the external environment.This kind of technology has given mobile robots the ability to autonomously locate and construct maps in unknown environments.The map created by the previous visual SLAM method is a collection of all landmark points,and determining the location of these landmark points represents the completion of the mapping task.This kind of sparse landmark map can be used to locate mobile robots,but because the spatial structure couldn't be infered from simple landmark and the system won't know whether there are obstacles or not,sparse map couldn't be used to complete the work that needs dense map to complete,such as navigation and obstacle avoidance.Aiming at the urgent need for mobile robots to require dense maps for large-scale localization and navigation,this paper proposes a real-time 3D reconstruction method.The specific work content of this article is as follows:Firstly,accurate pose estimation by visual-inertial odometry is obtained in this paper.Aiming at the problem of data fusion between camera and IMU observations,the preprocessing of stereo camera and IMU observations is studied.By pre-integrating the IMU,repeated integration is avoided when the state quantity changes.Through the loosely-coupled initialization method,the visual measurement is aligned with the IMU measurement,and a relatively accurate initial estimate is obtained.Subsequently,a tightly-coupled stereo visual-inertial odometry based on a sliding window is designed,and the real-time performance of the system was improved through using key-frame and marginalization while ensuring the accuracy of pose estimation.Secondly,a loop correction method is designed based on place recognition and pose graph for the problem of accumulated error in large-scale localization.Based on the pose graph output by the odometry,combined with place recognition,pose graph optimization and other methods,the loop correction of the pose trajectory is completed.The method can help the system to eliminate the accumulated error caused by long-term operation,thus ensuring the accuracy of the system.Then,a real-time 3D reconstruction method is designed based on superpixel segmentation for the problem of real-time dense mapping.The extracted superpixels are used to model the surfel in the system,so that the system can incorporate low-quality depth maps,ensuring the real-time performance of the system.Through local map fusion and map deformation,the established map can be globally consistent in real-time.Abundant prior information can be provided by this kind of reconstruction map for the following tasks such as navigation and operation of mobile robot.Finally,an experimental study of a visual 3D reconstruction system is carried,and the proposed algorithm is compared with a variety of advanced open source algorithms by using real world datasets.Experimental results show that the 3D reconstruction method proposed in this paper has high localization and mapping accuracy.High robustness in large-scale scenes is still maintained and real-time performance is ensured while only using CPU.