Abstract:

Visual SLAM（Simultaneous Localization And Mapping） is the core technology in the field of mobile robots. The traditional visual SLAM cannot be applied to highly dynamic scenes and the map lacks semantic information. This paper proposes a semantic SLAM method for dynamic environment. Firstly, the convolution neural network is used to detect the area of dynamic objects on the image. Secondly, it extracts feature points on the image and removes feature points in the dynamic area and then computes camera pose using static feature points. Finally, the dynamic objects’ map points are removed for building a global semantic map without interference from dynamic objects. The proposed method is tested on TUM datasets, and results show that the proposed method can improve the accuracy of pose estimation by 88.3% in dynamic scenes and construct semantic map without interference from dynamic objects.

Key words: visual Simultaneous Localization And Mapping（SLAM）, dynamic scene, object detection, semantic segmentation, semantic map

摘要：

视觉SLAM（Simultaneous Localization And Mapping，同时定位与建图）是移动机器人领域的核心技术，传统视觉SLAM还难以适用于高动态场景并且地图中缺少语义信息。提出一种动态环境语义SLAM方法，用深度学习网络对图像进行目标检测，检测动态目标所在区域，对图像进行特征提取并剔除动态物体所在区域的特征点，利用静态的特征点进行位姿计算，对关键帧进行语义分割，在构建语义地图时滤除动态物体的地图点构建出无动态物体干扰的语义地图。在TUM数据集上进行实验，结果显示该方法在动态环境下可以提升88.3%位姿估计精度，并且可同时构建出无动态物体干扰的语义地图。

关键词: 视觉同时定位与建图（SLAM）, 动态场景, 目标检测, 语义分割, 语义地图