Create synthetic LiDAR point clouds

Question

I'd like to create synthetic training data for DL models for segmentation and classification in point clouds. The ground truth / real data comprise LiDAR point clouds. I scripted a simple mesh sampling model in python/open3d and I'm able to quickly transfer 3D scenes to point clouds (see fig 1), but I need to include certain characteristics of LiDAR sensors.

Blensor ( https://www.blensor.org/ ) works the way I need it (fig 2), but I don't want to use blender atm. Also the results don't have a sufficient quality for my use case.

In the first step I'd just like to cut off the points, which are not reachable by a certain position of a LiDAR sensor, mainly to create the "shadows", which are important to make the training data more realistic. Do you have any suggestions for a simple and fast workaround? My point cloud is saved in a pandas dataframe including x,y,z and nx,ny,nz values.

Thx in advance, reiti

Answer 1

If your 3D scene can be described in the form of distance functions (essentially consisting of a range of simple geometric shapes opposed to point cloud data) you may be good to go with an easily modified ray tracing algortihm that emulates a lidar sensor.

For each lidar "ray" (ie for every direction) you only need to save the first scene collision's xyz coordinates. This also gives you full freedom to match the original real world sensor properties (like angles and number of points).

How easy the calculation of the distance between scene and sensor-ray will be, depends on the scene you have set up and how it is represented. Sorry for not being able to provide you with a ready to use implemntation but this might give you some direction.