A Visual SLAM Algorithm Based on Image Semantic Segmentation in Dynamic Environment

Most Simultaneous Localization and Mapping methods assume that the environment is static, and the scenarios they can apply to are strictly limited by the static environment. In practical application, it is affected or even fails because of the moving objects in the field of vision. We propose a visual SLAM algorithm based on deep learning for dynamic environments. Our method combines instance segmentation network and optical flow method, focusing on how to reduce the impact of dynamic objects in visual-based SLAM, so as to improve the localization accuracy in dynamic environment. It is composed of four modules: semantic segmentation module, dynamic point detection module, dynamic point removal module and feature-based SLAM framework. First, we use Mask RCNN to perform pixel-level semantic segmentation on the input image. Then, the Pyramid Lucas-Kanade optical flow is used to track and match the interframe feature points, and the dynamic point detection is carried out. In the dynamic point removal module, a dynamic point removal algorithm is proposed to deal with the feature points of dynamic objects, which greatly reduces the error of camera pose estimation caused by incorrect matching points. After that, the remaining static feature points are input into the feature-based SLAM framework for localization and mapping. We experimented on the open TUM dataset and compared it with three traditional SLAM algorithms in dynamic environments. Experimental results show that the new method has remarkable accuracy and improved robustness in complex dynamic environments.