Visual Slam in Dynamic Scenes Based on Object Tracking and Static Points Detection

Simultaneously Localization and Mapping (SLAM) plays a key role in tasks such as mobile robots navigation and path planning. How to achieve high localization accuracy in various scenarios is particularly important. This paper proposes a visual Semantic SLAM algorithm based on object tracking and static points detection, in order to eliminate the influence of dynamic objects on localization and mapping. This algorithm is improved on the framework of ORB-SLAM2. For continuously acquired input images, tracking algorithm is combined with the object detection to achieve the inter-frame correlation of objects in the scene. Then, epipolar geometry is used to detect static points on each object, and depth constraint is introduced to improve robustness. After excluding dynamic objects, the static points are sent to the tracking thread to achieve more accurate localization result. Finally, we record the pose of the dynamic objects for robots autonomous navigation in the future. Experiments on the public datasets TUM and KITTI show that in dynamic scenes, the proposed algorithm has reduced the relative index of absolute trajectory error (ATE) by more than 90% compared with ORB-SLAM2. Our system is also superior than DynaSLAM and DS-SLAM in most cases, which proves that the proposed algorithm can effectively improve the localization accuracy of visual SLAM in dynamic scenes.