???????????????????????????????????????????????? ????????????????????????????? 

Abstract

Visual Simultaneous Localization and Mapping (visual SLAM) enables robots to navigate in an unfamiliar environment. Most visual SLAM techniques can operate effectively under a static environment where objects are stationary. However, in practical applications, the environment often consists of moving objects, which causes visual SLAM to misidentify the sensor's position. In this paper, we propose an additional process to enhance the performance of visual SLAM for a dynamic environment. Our proposed visual SLAM system for dynamic circumstances, based on ORB-SLAM2, combines the capabilities of dynamic object detection to reduce the effect of dynamic objects. The system can detect moving objects by applying semantic segmentation and geometric method using both optical flow and epipolar constraint methods. The experiment was carried out on the Dynamic sequence of TUM RGB-D dataset and dataset from Realsense D435i camera. The experimental results show that the Absolute Trajectory Error (ATE) compared with standard visual SLAM baselines reduce up to 98.03% in TUM RGB-D dataset, and 83.87% in capturing video from Realsense D435i camera. It can fully demonstrate that the proposed object detection process can detect movable objects and reduce the impact of dynamic objects in visual SLAM. The average processing time is 0.2380 seconds per frame, therefore, the usage require processing on high-performance hardware to be used in real-time.

???????????????????????????????????????????????? (visual SLAM) ???????????????????????????????????????????????????????????????????????????? ??????????????????????????? ???????????????????????????????????????????????????????????????????????? ?????????????????????????????? ?????????????????????????????????????????????????????????????? ????????? visual SLAM ???????????????????????????????????????????? ????????????????????????????????????????????????????????????????????? visual SLAM ????????????????????????????????? (Dynamic environment) ??????????????????????? visual SLAM ??? ORB-SLAM2 ?????????????????????????????????????????????? (Dynamic object) ??????????? ???????????????????????????????????????????????????? ???? Semantic segmentation ??????????????????????????????????????? ????? (Optical flow) ??? ?????????????????? (Epipolar constraint) ??????????????????????????????????????????????????????????????????????????? RGB-D TUM ?????????? Dynamic objects ???????????????????????????????? Realsense D435i ??????????????????????????????????????????????????????????? (Absolute trajectory error) ??????????????????????????? ORB-SLAM2 ???????????????????? ?????????????????????????? RGB-D TUM ??????? 98.03 % ????????????????????????????????????????? Realsense D435i ????? ????????????????????? 83.87 % ??????????????????????????????????????????????????????????????????????????? visual SLAM ??? ???????????????????????????????????????? 0.2380 ???????????????? ????????????????????????????????????????????????????????????????????????????????????? (Real-time)