Adaptive LiDAR Sampling and Depth Completion Using Ensemble Variance.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Depth completion is crucial for autonomous vehicles, requiring accurate depth estimation from limited pixel measurements.
• Existing methods often rely on random or grid sampling, which can be inefficient for depth completion tasks.

Purpose of the Study:

• To develop a strategic and dynamic pixel sampling method for depth completion that maximizes accuracy.
• To enhance depth estimation for autonomous vehicles in various conditions (day/night) using programmable LiDAR.

Main Methods:

• An ensemble of predictors is used to determine pixel sampling probabilities, prioritizing pixels with high prediction variance.
• A multi-phase prediction approach minimizes redundant sampling of similar pixels.
• The ensemble method is compatible with various depth completion algorithms, including neural networks and Random Forests.

Main Results:

• The proposed ensemble-based sampling strategy significantly improves depth estimation accuracy compared to random or grid sampling.
• Achieved 4-10x reduction in required measurements for equivalent accuracy on the KITTI dataset.
• Implemented successfully with both state-of-the-art neural networks and a novel Random Forest-based learner.

Conclusions:

• The dynamic, ensemble-driven sampling method offers a highly efficient approach to depth completion for autonomous driving.
• This strategy maximizes the utility of limited depth measurements, advancing the field of real-time perception for autonomous systems.