Efficient 3D Lidar Odometry Based on Planar Patches.

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• 3D Lidars are crucial for autonomous vehicles, generating range images.
• Efficient real-time odometry is essential for accurate sensor trajectory estimation.
• Existing Iterative Closest Point (ICP) algorithms face computational challenges.

Purpose of the Study:

• To develop a novel, efficient real-time 3D lidar odometry algorithm.
• To leverage the ordered format of range images and exploit planar surfaces.
• To improve localization accuracy and computational speed compared to state-of-the-art methods.

Main Methods:

• Utilizes a 'flatness image' to identify planar patches in lidar range images.
• Extracts keypoints associated with planar patches for ego-motion computation.
• Imposes coplanarity constraints between point-plane pairs with iterative correspondence updates.

Main Results:

• The proposed method achieves 5x speedup over multi-threaded GICP with higher accuracy.
• A second version further reduces drift and computation time, running at 10-20 Hz on a CPU.
• Demonstrates superior performance compared to state-of-the-art ICP algorithms.

Conclusions:

• The novel flatness-based approach offers a significant advancement in real-time 3D lidar odometry.
• The algorithm provides a computationally efficient and accurate solution for autonomous vehicle localization.
• This method enables robust sensor trajectory estimation crucial for safe navigation.