Synthetic Training of Neural Networks for Semantic Segmentation of LiDAR Point Clouds

Abstract

Apron operations must ensure both high utilization of given capacity and safe aircraft operations even under degraded environmental conditions, such as low visibility. An appropriate sensor environment could support controllers, where deep learning models will ensure that the observed objects are classified correctly. The fundamental challenge is that these models require a large amount of data to be trained. Therefore, we have developed a virtual airport to generate the required training and validation data at any time and for any operational scenario (ground truth). We apply our concept of a virtual airport and sensor environment at Singapore Changi Airport implementing a synthetic LiDAR sensor. With the help of different data sources and own models, a multitude of 3D scenes can be generated which correspond to the real operational environment. From these scenes, a point cloud is extracted according to the specifications of the LiDAR sensor, which is already labeled by the underlying model and serves as input for PointNet++ for segmentation and classification. We show that the training of a classifier based on artificial input data is a promising approach, which covers relevant aspects of the real system and can therefore be easily applied in (augmented) tower environments.