A Visual SLAM Algorithm Based on Dual Attention Semantic Segmentation

Aiming at the problem that visual SLAM algorithms have difficulty in rejecting occluded dynamic objects in dynamic environments, proposed a Dual Attention semantic Segmentation algorithm for visual SLAM (DAS-SLAM). The algorithm enhances feature extraction by improving the deep backbone network, and introduces spatial attention mechanism and channel attention mechanism to build the scene semantic dependency of features to accurately identify occluded objects. The region marked by the instance segmentation uses a combination of multi-view geometry and dynamic thresholding to determine the motion state of objects in the region, eliminating the influence of dynamic targets in the scene on SLAM localization accuracy. The results show that the rootmean-square error of the trajectory of the algorithm is reduced by 36% compared with the DS-SLAM algorithm, which effectively improves the accuracy of camera pose estimation and map construction and demonstrates more robustness.

I. Introduction

With the rapid development of technology, robot navigation technology has been fully developed ˈ and Simultaneous Localization and Mapping (SLAM) technology is a major research hotspot in the field of computer vision and robotics. Its core lies in the utilization of robots in unknown environments. Self-equipped with sensors to estimate its own position in real time and build an environmental map [1]. Depending on the sensor used by the robot, there are two types of SLAM operation: laser SLAM and visual SLAM (vSLAM). The disadvantages of laser generators in acquiring colour and texture limit the application of laser SLAM, making the use of cameras to acquire environmental information the mainstream of SLAM. [2].