SDG-SLAM: an improved visual semantic SLAM algorithm towards dynamic environments

Visual simultaneous localization and mapping (SLAM) in dynamic scenarios is vulnerable to dynamic object factors, which can cause inaccurate pose estimation and limited robustness. To tackle this problem, we present a SLAM algorithm named SDG-SLAM, which combines improved visual semantic and geometric constraints to reject outliers and enhance performance. Firstly, the algorithm utilizes the neural network YOLOv5 (the fifth version of You Only Look Once) to obtain object’s prior semantics in the environment. In the meantime, the ORB feature points of keyframes are acquired, and the sparse optical flow tracking and matching are performed on non-key frames Secondly, adaptive initial dynamic point detection is implemented based on the epipolar constraints. Subsequently, the method of motion constraint posterior combined with semantic prior was used to screen and eliminate dynamic feature points. Finally, all static feature points from the environment are utilized, and these feature points are participated in the camera’s pose estimation. Substantial experiments performed on challenging public dynamic scenario datasets and outdoor dynamic scenarios demonstrate that SDG-SLAM significantly improves pose accuracy in scenes containing objects with different motion states. Specifically, compared to ORB-SLAM3, SDG-SLAM achieves more than 87% improvement in localization accuracy for high dynamic datasets and over 85% improvement for outdoor dynamic environments.