Object Mobility classification based Visual SLAM in Dynamic Environments

Most of the existed visual odometry methods cannot work in dynamic environments since the dynamic objects lead to wrong uncertain feature associations. In this paper, we involved a learning-based object classification front end to recognize and remove the dynamic object, and thereby ensure our ego-motion estimator’s robustness in high dynamic environments. Moreover, we newly classify the environmental objects into static, movable and dynamic three classes. This processing not only enables the ego-motion estimation in the dynamic environment but also leads to clean and complete map-ping results. The experimental results indicate that the proposed method outperformed the other state-of-the-art SLAM solutions in both dynamic and static indoor environments.