Vessel drift detection from point clouds

An example of a point cloud, colored by intensity, sourced from the laser at one side of the berth

Frontier Automation offers a 3DPortGuard product where they use lasers to detect and control a vessel movement and drift. They wished to know if machine learning could ingest point-clouds and provide a more robust estimate of a vessel’s drift.

The latency and throughput requirements of this application where challenging, and right at the limits of published work as of 2018. We tried various models including PointNet++, KDNet, and FlexConv. The result was compared to conventional algorithms such as the ones in the Point Cloud Library.

Under certain circumstances, it can deliver additional robustness to some of the situations that real-world deployment can throw at you, such as completely new vessels or vessels with very few unique features.

We think that machine learning, particularly un-gridded approaches, has a lot of potential for processing point clouds, particularly when you need results faster than a human can provide, with more robustness, or scale at a lower cost than human labeling.

An example of a similar coal loading and offloading facility