Towards Automated Apron Operations - Training of Neural Networks for Semantic Segmentation Using Synthetic LiDAR Sensors

For safe operations at the airport apron, controllers are supported by an appropriate sensor environment. Deep learning models could improve the classification of observed objects, but these models require a large amount of data to be trained. Therefore, we developed a virtual airport environment to generate the required training and validation data for any operational scenario. A synthetic LiDAR sensor is implemented in this environment and applied at Singapore Changi Airport. Using different data sources, the airport infrastructure and objects are modeled and a multitude of 3D scenes are generated. From these scenes, a point cloud is extracted from the LiDAR sensor feedback. This point cloud is already labeled by the underlying models (ground truth) and serves as input for PointNet++ to be trained for efficient segmentation and classification. We show that the training with synthetic input data is a promising approach even assuming degradation of the sensor feedback.