lec1-4-work-with-pcd.md

Lesson 1-4 Working with Real PCD

This lesson will apply the segmentation and clustering techniques to real point cloud data from a self-driving car.

• Real-world challenges for Lidar

  • Environmental conditions: heavy rains, stand storms, and anything reflects/scatters laser
  • Ghost objects caused by highly reflective surfaces: water sprays from other cars, etc

• Downsampling

  • Sending out the entire point cloud data is too heavy for the vehicle internal network
  • Use stixels
    • Stixels are segments which represent sensor data in a compact fashion while retaining the underlying semantic and geometric properties
    • An example of stixels (Ref). Use height, width, and number of rectangles to represent objects, instead of individual points.

I. Downsampling

Voxel grid filtering

Voxel grid filtering will create a cubic grid, thinking about a voxel grid as a 3D tiny box, over the input point cloud data points. It will downsample the cloud by only leaving a single point per voxel grid, so the larger the grid length the lower the resolution of the point cloud.

• Two options for selecting that representative point per voxel
  • Centroid of the point distribution, i.e. Spatial averaging (slower but more accurate)
  • Geometrical Center of the voxel

Region of interest (ROI) filtering

ROI-based filtering defines a boxed region and any points outside the box will be removed.

• Two regions interested
  • An adequate mount of distance in front of and at back of the car, with an approximate width of the road
  • Outside the rooftop area