Visual SLAM algorithm based on semantic information and geometric consistency1

Abstract

Localization through intricate traffic scenes poses challenges due to their dynamic, light-variable, and low-textured nature. Existing visual Simultaneous Localization and Mapping (SLAM) methods, which are based on static and texture-rich assumptions, struggle with drift and tracking failures in such complex environments. To address this, we propose a visual SLAM algorithm based on semantic information and geometric consistency in order to solve the above issues and further realize autonomous driving applications in road environments. In dynamic traffic scenes, we employ an object detection network to identify moving objects and further classify them based on geometric consistency as dynamic objects or potential dynamic objects. This method permits us to preserve more reliable static feature points. In low-texture environments, we propose a method that employs key object categories and geometric parameters of static scene objects for object matching between consecutive frames, effectively resolving the problem of tracking failure in such scenarios. We conducted experiments on the KITTI and ApolloScape datasets for autonomous driving and compared them to current representative algorithms. The results indicate that in the dynamic environment of the KITTI dataset, our algorithm improves the compared metrics by an average of 29.68%. In the static environment of the KITTI dataset, our algorithm’s performance is comparable to that of the other compared algorithms. In the complex traffic scenario R11R003 from the ApolloScape dataset, our algorithm improves the compared metrics by an average of 25.27%. These results establish the algorithm’s exceptional localization accuracy in dynamic environments and its robust localization capabilities in environments with low texture. It provides development and support for the implementation of autonomous driving technology applications.