Efficient Semantic-Aware TSDF Mapping with Adaptive Resolutions

Mapping unknown environments based on vision sensors is a key step for robot planning. Existing methods usually adopt a fixed resolution to store the occupancy information. A high resolution consumes excessive computational resources while a low one results in unbearable information loss of specific user-interested regions that require precise representation. To address these challenges, we develop an efficient semantic-aware Truncated Signed Distance Function (TSDF) mapping system with adaptive resolutions. Our map representation consists of multiple layers with different voxel sizes. Meanwhile, with the assistance of semantic segmentation or object detection models, we identify key objects in the scene and use higher resolutions for their reconstruction. Semantic information can guide the selection of map resolution and also potentially provide valuable insights for more complex user-defined navigation tasks. By leveraging our lightweight implementation, our system achieves real-time computation and less memory consumption than existing frameworks.