泡泡一分钟：Semantic Monocular SLAM for Highly Dynamic Environments

Semantic Monocular SLAM for Highly Dynamic Environments

高动态环境的语义单目SLAM

Nikolas Brasch , Aljaz Bozic , Joe Lallemand , Federico Tombari

摘要 - 单目SLAM的最新进展已经实现了具有实时能力的系统，该系统在静态环境的假设下运行稳健，但在动态场景变化和运动的情况下失败，因为它们缺乏明确的动态异常值处理。我们提出了一种语义单目SLAM框架，旨在处理高度动态的环境，结合基于特征的方法和直接方法，以在具有挑战性的条件下实现稳健性。所提出的方法利用在显式概率模型内从场景提取的语义信息，这最大化了跟踪和建图两者依赖于相对于相机不呈现相对运动的那些场景部分的概率。我们在动态环境中显示更稳定的姿态估计，并且在Virtual KITTI和Synthia数据集上的静态序列上具有与现有技术相当的性能。

Abstract— Recent advances in monocular SLAM have en-abled real-time capable systems which run robustly under the assumption of a static environment, but fail in presence of dynamic scene changes and motion, since they lack an explicit dynamic outlier handling. We propose a semantic monocular SLAM framework designed to deal with highly dynamic en-vironments, combining feature-based and direct approaches to achieve robustness under challenging conditions. The proposed approach exploits semantic information extracted from the scene within an explicit probabilistic model, which maximizes the probability for both tracking and mapping to rely on those scene parts that do not present a relative motion with respect to the camera. We show more stable pose estimation in dynamic environments and comparable performance to the state of the art on static sequences on the Virtual KITTI and Synthia datasets.

这里提出了用于高动态环境的单目SLAM方法，其使用基于CNN预测的语义先验信息的联合概率模型来模拟动态异常值。