Visual SLAM Method Based on Semantic Segmentation in Dynamic Scenes

Abstract

A semantic visual SLAM algorithm based on an improved semantic segmentation network DeepLabv3plus and multiview geometry is designed to address the issues of poor robustness and susceptibility to interference from dynamic objects in visual Synchronous Localization And Map (SLAM) construction in dynamic scenes. Based on the semantic segmentation network DeepLabv3plus, a lightweight convolutional network MobileNetV2 is used for feature extraction, and depthwise separable convolutions are used instead of standard convolutions in the Atrous Spatial Pyramid Pooling (ASPP) module. Simultaneously, an attention mechanism is introduced to propose an improved semantic segmentation network DeepLabv3plus. Combining the improved semantic segmentation network DeepLabv3plus with multiview geometry, a dynamic point detection method is proposed to enhance the robustness of visual SLAM in dynamic scenes. On this basis, a three-dimensional semantic static map containing both semantic and geometric information is constructed. The experimental results on the TUM dataset demonstrate that compared with ORB-SLAM2, the highest Root Mean Square Error (RMSE) and Standard Deviation (SD) values increased by more than 98% and 97%, respectively.