GMP-SLAM: A real-time RGB-D SLAM in Dynamic Environments using GPU Dynamic Points Detection Method

摘要

Simultaneous Localization and Mapping (SLAM) is a fundamental technology for robotics. Vision-based SLAM has been developed for many years, but it is still difficult for SLAM system to handle dynamic environments. In this paper, we present GMP-SLAM, a realtime RGB-D SLAM system for highly dynamic environments with the help of GPU Grid Map Projection— a GPU dynamic points detection method we design. SLAM is a time-sensitive system for robotics, and it is hard to reach real time, especially in dynamic environments because it is necessary to track moving objects and it takes a lot of time. GMP-SLAM is based on ORB-SLAM2, which is one of the best feature-based SLAM frameworks and can reach real time just in CPU. But ORB-SLAM2 cannot handle highly dynamic environments very well, and most work focus on tracking moving objects with a neural network, which cannot reach real time even with the help of GPU. To solve real-time problem, we propose an all-in-parallel dynamic points detection framework for visual simultaneous localization and mapping (VSLAM) in dynamic environments based on 3D occupancy grid maps. Our SLAM system can provide not only higher trajectory accuracy but also a 3D grid map for navigation. We test our SLAM system in our real-world datasets we record and get higher trajectory accuracy than ORB-SLAM2. At the same time, our system can run nearly in 20Hz, which is much better than existing VSLAM framework in dynamic environments.