3D Convolutional Neural Networks for Tree Detection using Automatically Annotated LiDAR data

Abstract

Methods In order to identify trees in LiDAR scans, ground points are first identified and filtered using a Progressive Morphological Filter. This filtered scan is then voxelized in a sparse 3D hierarchical data structure, VOLA (Byrne et al., 2017), in order to reduce the input resolution. A 2 bits per voxel approach is used to encode additional information such as colour, intensity and number of returns information.