Semantic Evidential Grid Mapping Using Monocular and Stereo Cameras.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Accurate local traffic scene estimation is crucial for automated vehicle software development.
• Existing grid mapping methods often rely solely on range sensors (Lidar, Radar) and lack semantic information.
• Integrating vision-based sensors offers redundancy and richer scene data.

Purpose of the Study:

• To develop a semantic evidential grid mapping pipeline incorporating vision sensor data.
• To enable fusion of vision data with existing range sensor data for comprehensive scene representation.
• To explicitly model uncertainties within the evidential model for robust estimation.

Main Methods:

• A novel semantic evidential grid mapping pipeline processing monocular and stereo vision data.
• Incorporation of disparity- or depth-based ground surface estimation for accurate mapping.
• Explicit modeling of uncertainties in the evidential model.

Main Results:

• Accurate and dense semantic grid maps generated from vision data.
• Successful fusion of vision-based semantic information with potential range sensor data.
• Demonstrated superiority over existing semantic grid mapping approaches in quantitative evaluations.

Conclusions:

• The proposed pipeline effectively enhances traffic scene understanding for automated vehicles using vision sensors.
• Explicit uncertainty modeling improves the robustness of semantic grid maps.
• The approach provides a foundation for sensor-diverse and semantically rich environmental representations.