Random Forest Machine Learning for classification of LIDAR data

In this presentation we explain how to classify objects detected in LiDAR data using a Random Forest and other machine learning classification methods. The focus of the talk is explaining Random Forest and how to optimize the parameters to improve accuracy.

LiDAR (Light Detection And Ranging) data is collected by firing a laser many times with the hardware to determine the direction and distance to the surface hit by the laser. In our data, the LiDAR is on board an aircraft, resulting in x-y-z locations for about 18 million points on objects hit by the laser. We overlay a grid with about a million squares over the ground, and bin the x-y-z points into the grid squares, so that each square contains 10-20 points detected by the laser. Our goal is to determine the object in each grid square (building, vehicle, road, tree, etc.), and our features are statistics for the heights of points in each square (mean height, max height, min height, and standard deviation of heights within each square, as well as statistics over neighboring squares).

Our object classes include buildings, building rooftop structures, forest trees, landscape trees, landscape bushes, cars, light posts of varying sizes, fences, paved surfaces, and grass. Our classification method of choice is a Random Forest, but we also investigate other machine learning methods including K-Nearest Neighbors and Linear Discriminant Analysis. We evaluate the effectiveness of the algorithms for accuracy, required training sample size, and runtime.

Ground truth was created in Gimp, classification processing was done in Python, and the fly-through visualization at the end was done using Blender.

The full paper is available here:
https://www.researchgate.net/publicat...

More information about LiDAR can be found here:
   • What is Lidar?  How does Lidar work? ...