A Visual SLAM Algorithm Based on Dynamic Feature Point Filtering

Abstract

High-precision positioning is the basis of robotics and driverless technology. Using the Visual Simultaneous Localization and Mapping (VSLAM) algorithm for positioning can effectively reduce the dependence on satellite positioning signals. However, there are a large number of moving objects in the road environment seriously affecting the positioning effect of VSLAM algorithm. YOLOv4-Tiny is introduced to identify dynamic objects, and Meanshift algorithm is used to identify dynamic feature points to form YMS(YOLOv4-Tiny, Meanshift) dynamic feature point filtering algorithm. The YMS-SLAM algorithm is proposed on the basis of VINS-Fusion, and the dynamic feature points are filtered out by the YMS algorithm to eliminate the adverse effects of dynamic objects on VSLAM. The experimental results in 05 sequence of the odometry data of the KITTI dataset and real road environment show that the root mean square error of YMS-SLAM is reduced by 21.0% and 61.6% respectively compared with VINS-Fusion, which proves that the proposed algorithm has stronger robustness and higher positioning accuracy in the road environment with dynamic objects.