ALARM: Adversarially-learned Labels using Activity and Reward Models

Technological advances in navigation and positioning, along with expanding wireless infrastructure and remote sensing technologies, have resulted in explosive growth of available trajectory data from a variety of moving objects, such as people, cars, ships, and animals. Trajectory data have numerous commercial applications, e.g., location-based services, travel forecasting, health monitoring, disease spread forecasting, land use analysis, urban planning, and robotics. When high-resolution data are not available, intelligence analysts use target trajectories to perform behavior-based object classification and activity prediction. To deal with large amounts of trajectory data, analysts must rely on automated algorithms to perform trajectory-based analysis and reasoning. However, existing analytical applications for trajectory analysis are very limited. Supervised trajectory classification models require large amounts of labeled data or location metadata that do not exist in many military applications. Many unsupervised algorithms exist to perform trajectory similarity measurement, clustering, outlier detection, and motion prediction. However, these methods are sensitive to data noise (such as track breaks and irregular time sampling), do not generalize well to changing terrains and targets, and are unable to produce automated characterization of the motion data. Most importantly, state-of-the-art trajectory mining algorithms do not explain how and why the motions were generated, limiting their utility in GEOINT applications when data are unlabeled, noisy, and do not contain contextual layers.  \n\n To address the gaps described above, Aptima Inc. will develop a system that analyzes spatiotemporal trajectory data to produce Adversarially-learned Labels using Activity and Reward Models (ALARM). ALARM will generate the semantic labels of moving targets in the areas of interest by learning generative behavior models that maximally explain observed trajectories. ALARM will be able to produce meaningful and generalizable motion semantics because our algorithms disentangle and learn the intents and behaviors of moving entities from the movement constraints, thus capturing statistically consistent causes of behaviors generalized across multiple trajectories. ALARM will learn movers’ intent as the reward functions that explain why the trajectories occurred, while behaviors are learned as conditional action probability distributions specifying how trajectories could be generated. The “generative” nature of learned behavior models will enable three critical analytical tasks - explanation, classification, and prediction - because analysts not only would understand what has already occurred, but also be able to reason about and predict how the targets may modify their behaviors in different contexts and under changing goals.