Abstract. Point cloud data from aerial LiDAR scan (“ALS”) are used for object detection and classification of energy industry facilities and assets. It is advantageous to be able to carry out point cloud classifications in near real time on secure hardware at the survey location and to be able to rapidly train the model on custom object classes. Such requirements create the need for efficient deep learning architectures which produce accurate predictions with low computational cost and time. This research presents a solution using Modified Point Voxel CNN (“MPVCNN”) which consists of feature-level fusion between voxel and point features for local feature extraction. In doing so, this architecture circumvents indexing operations and GPU memory limitations. The MPVCNN developed in this research was trialled using dense DALES datasets. Additionally, Aerial LiDAR scan datasets typically suffer from a class imbalance for rare objects and those which are physically small or thin-shaped, relative to other object classes. This research explores how a second classification pass can be used to improve the initial classification prediction for such imbalanced object classes, by using predicted class labels as a criterion to group points which are semantically homogeneous in computing geometric features. This paper demonstrates that the MPVCNN architecture is capable of high accuracy (>0.9 F1-score and OA) classifications, with short training times (approximately 1 hour), on dense ALS datasets using standard hardware (e.g. 8GB GPU).
Othman et al. (Wed,) studied this question.