3D LiDAR Process – 3D point cloud

LiDAR L4B 2

In autonomous driving applications, the type of LiDARs used is mostly 3D LiDARs. Such sensors use multiple rotating laser beams propagating from the sensor with different inclination angles in order to provide a 3D point cloud representation of the surroundings. 

Filtering

Filtering data can be done using many different methods, depending on the required effects. Some examples are Directed filtering of a specific point/plane/region in the point cloud using prior knowledge with a Conditional Filter, or if a shape is known a priori, a Parametric Model Filter. Removal of noisy data around objects using a statistical outlier removal, down-sampling a high-density 3D LiDAR point cloud into voxels using the Voxel Grid approach, and removal of the ground plane using the RANSAC algorithm. 

Segmentation 

The human eye has naturally evolved to detect objects and shapes from noisy backgrounds efficiently. It can look at a noisy point-cloud image and point out any objects or shapes. However, this process is not trivial to program a machine to do. The following are a few examples of point cloud segmentation algorithms. 

We can find a surface plane that fits the point cloud (e.g., the ground vs. the face of a nearby building) using the RANSAC algorithm. The segmentation process can also be used given prior knowledge of a different shape (e.g., a cylinder). 

The Region Growing algorithm merges points close enough in terms of a smoothness constraint, thereby creating clusters that belong to an object surface. 

To improve the point cloud processing speed by a divide-and-conquer strategy, we can segment points in the point cloud based on their distance to some central locations (Euclidean Clustering). 

Using Voxelization to identify object features 

Much in the same way that blurring improves an algorithm’s ability to detect objects in a noisy image, our system can be built to identify the salient features in point cloud data. In 2D image analysis, over-segmentation of an image and grouping regions with similar pixels (known as superpixels) reduces the number of regions that must be considered later by more computationally expensive algorithms. For 3D LiDAR data, over-sampled and noisy 3D point cloud data are clustered into supervoxels (the 3D analog of superpixels). 

Conclusion 

Point cloud data processing is a crucial aspect of any LiDAR sensor system. However, much more goes into building an automotive LiDAR product, and L4B Software provides full product lifecycle management for the development and integration of automotive systems, including cameras, LiDAR, and Radar. 

New call-to-action