Research On Semantic SLAM For Dynamic Scenarios

Simultaneous Localization and Mapping(SLAM)refers to the process of a robot building a map of an unknown environment while simultaneously determining its own location using its sensors.Cameras are favored in the SLAM field due to their affordability and ability to capture rich information.However,early SLAM algorithms assumed that the robot was operating in a static environment,which is not the case in real-world scenarios where dynamic objects are present.The existence of dynamic objects can greatly affect the performance of the robot’s functions,such as reducing localization accuracy and inaccurate mapping.This paper focuses on the study of visual SLAM algorithms in dynamic indoor environments.Based on the traditional SLAM algorithm framework,improvements are made to reduce the impact of dynamic objects,improve the algorithm’s localization accuracy,and construct more accurate maps.The main contributions of this paper are as follows:(1)To address the impact of dynamic objects on SLAM algorithms,semantic segmentation is used to segment images,and optical flow is used to track feature points.Dynamic points are then identified to obtain a segmented image with semantic information.When the number of dynamic feature points in a certain region of the segmented image exceeds a predetermined value,all feature points in that region are discarded to improve the performance of the SLAM algorithm.(2)Based on the ORB-SLAM2 algorithm framework,the algorithm from(1)is incorporated into the framework.Dynamic feature points are removed,and static points are used for feature matching to improve the accuracy of the tracking thread.The algorithm is optimized using the pose graph,and the bag-of-words model is used to determine if the system has achieved loop closure,thus completing the entire SLAM algorithm process.(3)The overall SLAM algorithm framework is built,with an additional map construction thread added to the ORB-SLAM2 framework.The surrounding static environment is mapped after removing dynamic object information,and the semantic information segmented by the segmentation thread is then fused into the map to construct a semantic map.(4)The performance of this paper’s algorithm is compared with that of the ORB-SLAM2 algorithm and other excellent algorithms for localization accuracy and octree map construction using the TUM dataset.Experiments are also conducted using the Kinect 2.0 camera in real-world scenarios to verify the real-time and effectiveness of this algorithm.