Occupancy Grid Map Construction Based on Semantic Segmentation and a Priori Knowledge

Navigational map is a prerequisite for automatic guided vehicle. The traditional feature-based visual Simultaneous Localization and Mapping (vSLAM) systems extract sparse points to generate a map that cannot be used for navigation or path planning. Generally, dense depth estimation based on multi-view geometry or Convolutional Neural Network (CNN) is a typical Solution for vSLAM systems to construct a navigational map. However, depth estimation is sometimes inaccurate in low-texture or reflective regions and difficult to evaluate errors in practice. To improve these problems, we propose a solution named Semantics-guided Structure Reconstruction Mapping (SSR-Mapping) that utilizes a stereo camera to construct an indoor navigation map avoiding dense depth estimation. The key aspects of SSR-Mapping are semantic segmentation, priori knowledge of indoor structure features, and visibility constraint. A post-process method is also proposed to correct navigation map reconstruction errors resulting from some inaccurate semantic segmentation. Experiments are carried out to compare SSR-Mapping with the systems using dense depth estimation. The results validate the feasibility and show the promising performance of SSR-Mapping.