An Improved Pose Estimation Method in Dynamic Scene Based on Hierarchical Masking and Point Features Classification

Compared with laser SLAM, visual SLAM solutions have the advantages of low cost and rich collected environmental information, but the accuracy of their estimated pose and building maps can also be seriously disturbed by dynamic objects in the camera scene. For the purpose of eliminating the influence of dynamic objects, this paper proposes an improved pose estimation method based on ORB-SLAM2 and combined with the object detection and instance segmentation network. A special mechanism of mask generation is introduced, which takes the result of detection as a precondition, and adaptively adjusts the intervention of pixel-level segmentation mask, so as to reasonably balance segmentation accuracy and computational overhead. Then geometric constraints are adopted to identify the motion state of point features within the mask above. Finally, the static point features filtered twice are used to estimate the accurate pose of the camera. Experiments on the TUM datasets show that the proposed algorithm has significantly improved the trajectory accuracy and the anti-drift ability compared with some typical existing algorithms in test sequences containing dynamic objects, which proves the effectiveness of the above method.