Reproduce “Recognizing and Tracking High-Level, Human-Meaningful Navigation Features of Occupancy Grid Maps” on OpenResearch

About this paper

Payam Nikdel, Richard Vaughan

Researchers at Simon Fraser University developed a system for robots to recognize and track high-level human-meaningful navigation features, such as open rooms, closed rooms, and corridors, from occupancy grid maps. The system utilizes a Convolutional Neural Network (CNN) combined with a temporal tracking module to provide stable real-time semantic understanding, achieving F1 scores up to 0.92 for open rooms in simulated environments and verbally communicating these features for human-robot interaction.

Key insights

• The temporal tracking module significantly improves F1 scores across all detected feature classes and input modalities, confirming its effectiveness in stabilizing predictions and enhancing overall accuracy.
• While fusing both direct LiDAR and SLAM-generated map data (Combined model) generally yielded the best performance in simulation, direct LiDAR data alone proved more robust in real-world environments due to inaccuracies and delays in SLAM output.
• Open rooms are consistently the easiest features to detect due to their distinct visual signature and effective data augmentation, while corridor detection remains more challenging across environments due to varied geometries.

Summary by alphaXiv.