Artificial Neural Networks for Individual Tracking and Characterization of Wake Vortices in LIDAR Measurements

Abstract

A vortex detection and characterization strategy in LiDAR measurements (Light Detection and Ranging), consisting of two different parts, is presented in this study. First, a two-stage detection pipeline is implemented combining the computer vision deep neural network YOLO and a regression convolutional neural network to detect vortices individually. An accuracy in the order of the instrument accuracy is achieved. Second, a new characterization method the so-called Projection Method is presented which has has the following important features. First, it is very accurate. Second, it yields valuable quantification of the accuracy of the results. Third, it is very fast and outperforms conventional methods by two to three orders of magnitude, hence it enables to process a high amount of data in a short time. Fourth, it provides much flexibility in choosing different models and compare the performance of the respective models. Fifth, it comes with a natural visualization of the underlying calculus. Sixth, it can be generalized to situations, where measurements provide a reduced and skewed image of the reality and certain structures or features have to be identified and characterized employing models.