Scene as Occupancy and Reconstruction: A Comprehensive Dataset for Unstructured Scene Understanding

Area of Science:

• Robotics and Artificial Intelligence
• Computer Vision
• Autonomous Systems

Background:

• Autonomous driving technology is advancing towards large-scale commercialization, with safety and comfort as key performance metrics.
• Current research often focuses on urban driving, neglecting unstructured scenes with irregular obstacles and road undulations.
• Existing datasets and studies for unstructured environments are scarce, limiting the development of robust autonomous systems.

Purpose of the Study:

• To introduce the world's first comprehensive benchmark dataset for perception in unstructured scenes.
• To facilitate the expansion of autonomous driving applications beyond urban environments.
• To enable research on improving safety and comfort in autonomous driving through enhanced scene understanding.

Main Methods:

• Development of a novel perception dataset specifically designed for unstructured scenes.
• Inclusion of detailed annotations for 3D semantic occupancy prediction to detect irregular obstacles.
• Inclusion of road surface elevation reconstruction to characterize road surface conditions.

Main Results:

• The dataset provides comprehensive annotations for 3D semantic occupancy prediction and road surface elevation reconstruction.
• Trajectory and speed planning information is included to link perception with planning.
• Experiments with state-of-the-art methods validate the dataset's effectiveness and highlight task challenges.

Conclusions:

• This dataset is a valuable resource for advancing autonomous driving technology in unstructured environments.
• It addresses the critical need for data in complex, off-road scenarios.
• The benchmark facilitates research into perception, planning, and decision-making interpretability for autonomous vehicles.