LiDAR Data Classification Using Convolutional Neural Network Based on PointNet Architecture

Convolutional neural networks are the state-of-the-art algorithm for object classification. Due to the various types of objects that are processed and to facilitate training, typical convolutional neural networks (CNNs) require data preprocessing like zero padding or 3D to 2D space projections and do not work with point cloud data. Light Detection and Ranging (LiDAR) is one of the main technologies used in self-driving cars and terrain mapping. Since LIDAR uses time of flight from laser beams to create a 3D map of the area, the generated data is a point cloud. In order to solve these problems, here we present an implementation of CNNs using a modified PointNet architecture. PointNet architecture is directly capable of taking a point cloud and running it on the classification algorithm, which is much more efficient than transforming the data before being fed to the network. In this study, we optimize the said convolutional neural network based on PointNet architecture. We train the model using LiDAR data taken in Westwood and tune its parameters accordingly to achieve close to state-of-the-art performance. As of now, in preliminary testing, the model achieves an 89.82% training accuracy. The goal is to further achieve a model that can be able to map external environments to aid driver-safety and autonomous navigation.