Geometry-Aware Deep LiDAR Odometry

This research summary is just one of many that are distributed weekly on the AI scholar newsletter. To start receiving the weekly newsletter, sign up here.

Often, odometry modules are model-based and focus on using RGB-D cameras and LiDAR. Although they have high performances, model-based methods face challenges including vulnerability to environmental disturbance and parameter selections. That said, there have been efforts towards learning-based odometry using LiDAR.

The learning-based approach using LiDAR with deep learning has been reviewed in numerous research works in the past where researchers have followed a supervised learning framework and faced the challenge of handling a dense point cloud into a deep neural network. Since previous approaches rely on supervised learning which requires the ground-truth with labeled sequences, researchers have now come up with an unsupervised methodology for deep LiDAR odometry.

Deep Learning Odometry Approach

DeepLo is a new approach that represents the first-ever unsupervised learning-based odometry for LiDAR. DeepLo incorporates Iterated Closest Point (ICP) technique into a deep-learning framework and can be trained by either using supervised or unsupervised approaches. It also integrates two loss functions that permit switching between the supervised and unsupervised learning depending on the ground-truth validity in the training phase. For effective unsupervised training and prediction, the researchers apply both a vertex and a normal map as inputs and use them for loss calculations.

DeepLo has been assessed using well-known odometry KITTI and Oxford RobotCar standard benchmark datasets. The new approach demonstrates improved performance and efficiency for achieving pose accuracy.

Potential Uses and Effects

Autonomous researchers and engineers, as well as the entire AI community, can use DeepLo for LiDAR point cloud regardless of the configuration or hardware type to achieve scalability and flexibility during model training.

DeepLo also provides them with the capability to attain effective simultaneous localization and mapping (SLAM) for a wide variety of applications such as autonomous cars, robots, 3D mapping and more.

Thanks for reading. Please comment, share and don’t forget to subscribe! Also, follow me on Twitter and Medium. Cheers!