Visual SLAM With MmWave Point Cloud In Dynamic Environmen

With the development of intelligent robots and autonomous/assisted driving,simultaneous localization and mapping(SLAM)technology has gradually become the key to the application of autonomous mobile robots and autonomous driving.The visual SLAM system uses low-cost cameras as the main sensors for localization and mapping,but the current visual SLAM systems based on static environment assumptions have decreased accuracy and robustness in dynamic environments.This article focuses on the localization and mapping requirements of robots in real dynamic environments,and studies how to improve the accuracy and robustness of visual SLAM systems in dynamic environments.This article uses 4D millimeter-wave radar to detect dynamic objects directly in the environment,and then dynamic areas marked in the image.Based on the ORB-SLAM3,RDyna SLAM system is proposed,which has added 4D radar dynamic target clustering,dynamic mask generation,and dynamic feature screening algorithms.RDyna SLAM improve accuracy and robustness of SLAM systems in high dynamic environments.The main work of this article is as follows:1)Models of camera and millimeter wave radars are analyzed,as well as the architecture and functions of the ORB-SLAM3 system is analyzed.Accuracy of the ORBSLAM3 system in dynamic and static environments in the TUM dataset are tested and compared,and the reasons for the decrease of accuracy in dynamic environments are studied.Finally,improvement methods for the ORB-SLAM3 system are proposed.2)A highly dynamic environmental dataset is proposed including visual and 4D millimeter-wave radar data.Firstly,a correlation model between the camera and radar was constructed,and the internal and external parameters of the correlation model were calibrated.Then experimental environment is divided into corridors,laboratories,offices,and outdoor scenes withing dynamic objects such as pedestrians,cars,bicycles.During data collection,the sensors are set to rest and dynamic objects are set to occupy the main part of the image to simulate a highly dynamic case.Finally,the a visual-4D radar dataset is proposed in a high dynamic environment.3)A pose estimation method for high dynamic environments is proposed by fusing millimeter wave radar point clouds.Firstly,this article uses doppler information from millimeter-wave radar to separate dynamic radar points from static background radar points.Then,the dynamic clusters are expanded after projected onto the image from radar points.Dynamic masks generated based on the distribution characteristics of the expanded clusters.Finally,in response to the interference of dynamic feature in pose estimation,a dynamic feature screening method was designed based on dynamic feature masks.A robust pose estimation method in high dynamic environments is proposed.4)Based on ORB-SLAM3,the RDyna SLAM system integrating millimeter wave radar is proposed and deployed on a robot platform.Finally,the accuracy of the SLAM system in high dynamic environments is tested in the dataset proposed in this paper.The results show that the absolute pose error of RDyna SLAM in high dynamic environments is reduced by an average of 69% compared to ORB-SLAM3 system,and 96% in some case.