Real-Time Visual Recognition of Ramp Hand Signals for UAS Ground Operations

Abstract

We describe the design and validation of a vision-based system that allows the dynamic identification of ramp signals performed by airport ground staff. This ramp signals’ recognizer increases the autonomy of unmanned vehicles and prevents errors caused by visual misinterpretations or lack of attention from the pilot of manned vehicles. This system is based on supervised machine learning techniques, developed with our own training dataset and two models. The first model is based on a pre-trained Convolutional Pose Machine followed by a classifier, for which we have evaluated two possibilities: A Random Forest and a Multi-Layer Perceptron based classifier. The second model is based on a single Convolutional Neural Network that classifies the gestures directly imported from real images. When experimentally tested, the first model proved to be more accurate and scalable than the second one. Its strength relies on a better capacity to extract information from the images and transform the domain of pixels into spatial vectors, which increases the robustness of the classification layer. The second model instead is more adequate for gestures’ identification in low visibility environments, such as during night operations, conditions in which the first model appeared to be more limited, segmenting the shape of the operator. Our results support the use of supervised learning and computer vision techniques for the correct identification and classification of ramp hand signals performed by airport marshallers.