LiDAR Machine learning
Overview
LiDAR which is an acronym for light detection and ranging is a method of remote sensing. This method consists of illuminating an object plane with a laser source and capturing the reflected light from that plane. The collected light would then be processed to render a 3-dimensional model of the object plane. The collected data is stored as a point cloud data set where each point represents a reflection from the object. Since the principle of detection in LiDAR in distances, it has been difficult to teach a computer how to combine points in space to create an object as well as distinguish different groups of data points.
New research by the Computer Science and Artificial Intelligence Lab (CSAIL) at MIT, developed a machine-learning algorithm that is capable of detecting objects using LiDAR. This is possible by creating a point cloud data representation of individual scans and identifying and recognizing what is present at each scan. Their method consists of creating a dynamic graph convolutional neural network. This method allows the user to classify individual objects within each scan. This allows the computer to take multiple scans and find hierarchical patterns of generic information that can be used by downstream tasks. The biggest issue with machine learning for LiDAR scanners is that the position of each object typically changes with each scan. To overcome this researchers at CSAIL, presented a method called deep closest point (DCP). This allows the computer to find similar edges and shapes within different scans and superimpose their features such that the computer can recognize and classify objects.
Related Links
Article: Deep learning with point clouds
Article: 3D object detection with Lidar