A Real-time Semantic Elevation Map Construction Method Based on Hierarchical Residual Network

Semantic elevation map integrates spatial geometric information with semantic attributes, providing multimodal perception and decision-making support for mobile robots. However, existing mapping methods face challenges when handling object-background boundary ambiguities in multi-object dynamic environments, leading to reduced accuracy in defining traversable areas. To address these challenges, we propose a real-time semantic elevation map construction method based on hierarchical residual network: 1) An improved YOLACT segmentation model with hierarchical residual connections is employed to generate image segmentation masks for stairs, pedestrians, and obstacles; 2) The 2D segmentation masks are fused with 3D point cloud features to compute environmental height information and determine semantic attributes, constructing the semantic elevation map. Experimental results demonstrate that the improved image segmentation module outperforms existing methods, and the semantic elevation map accurately estimates both the terrain height and semantic categories of the environment.