Classifying Airborne Targets With Motion-Based Recognition Techniques

This work demonstrates how motion-based recognition techniques can be used to classify different types of airborne targets with passive data from a stationary ground-based sensor. Using motion-based information for classification is common practice in applications such as lipreading, hand gesture interpretation, and gait recognition (walking vs. skipping vs. running). In all these scenarios, one is tasked with classifying a 3D motion event from a sequence of 2D images. The main contribution of this work lies in transitioning these ideas into the realm of airborne target classification. Specifically, we look at a scene consisting of synthetically generated trajectories for small birds, large birds, multi-rotor drones, and fixed-wing drones. The projected image plane motion of these different types of targets will vary due to their different objectives and distinct methods of achieving flight. We can parameterize these characteristic trajectories using a variety of different motion-based recognition techniques in order to generate motion-based features (such as speed, direction, and spatiotemporal curvature) that can be used for classification. Our results show the performance of two different motion-based classifiers and demonstrate how how their outputs can be combined with an ensemble method to take advantage of each of their strengths.