Author: Yoon, Barfoot & co
Year: 2021
- feature and uncertainty learning with dnn for LiDAR odometry
- sparse structure of probabilistic formulation
- no groundtruth is required and network trained online
- does not need the estimator to be differentiable
- learning objective takes into account uncertainty
- approximate posterior as a multivarite gaussian distribution
Unsupervized DL for lidar odometry:
- optimisation over a w frames window
- apply EM (with E step optimizing on posterior and M step optimizing on parameters) for learning
- KPConv: convolution layer for point cloud based on sherical kernel
- The network compute descriptor, measurement covariance matrix and detection score for each point
- Then a keypoint is computed for each voxel in a voxel grid using DNN output
- E step is basically factor graph optimization
- Hard thresholding on Mahalanobis distance for outlier rejection before M-Step
- Using covariance to evaluate the quality of a keypoint as well
Results:
- KITTI process LIDAR data to take distorsion into account
- Not running in real time because of KPConv layers