Anomaly Detection for Trajectory Data

Advances in sensors and the widespread use of devices with geo-referencing capabilities have led to a profusion of data on trajectories of objects. These trajectories are frequently associated with important activities such as monitoring the trajectories of endangered species in the ocean or coping with potential security threats by people or ships. Anticipating unexpected events in these trajectories is of key importance as it may allow for preventive actions to be put in place. The long-term objective of this research program is to develop data-driven modelling algorithms to help forecast anomalies in trajectory data.

Anomalies are by definition rare, which creates an imbalanced prediction task that is challenging for the modelling algorithms. While imbalance has been extensively studied for classification tasks, regression problems have received less attention, and our recent work has helped to fill this gap. With the current proposal we plan to extend these approaches to tasks, such as trajectory data, where observations are not independent, with clear temporal, spatial and/or spatiotemporal dependencies. Ignoring these dependencies may lead to unreliable models being obtained. A principal objective is thus to develop new modelling techniques for data that are imbalanced and exhibit dependencies, with a focus on moving objects in the ocean such as ships and animals.

Feature engineering is a critical element of complex forecasting tasks. A second main research theme of this program is the study and development of feature-engineering methods that capture important dynamic characteristics of trajectories with the goal of improving the accuracy of models. We will focus our studies on developing features that describe deviations from normal trajectories, e.g. difference of the velocity of a moving object to the average velocity of other objects with similar trajectories. Our research hypothesis is that these features will help with the goal of being able to anticipate critical events in trajectory data. A key aspect of our scientific approach will be the use of input from domain experts to guide informed feature design, which will be applied to real world trajectory data of fishing boats. This approach is motivated by the complexity of the Ocean ecosystems.

The significance of the program results from the importance of trajectory data for several key application domains. This program research will advance the state of the art on trajectory mining by addressing two relevant but still little explored issues: (i) feature engineering for this type of data; and (ii) forecasting anomalies on trajectories. In this research program we will focus on application areas with high socio-economic relevance related with the Ocean, with the main application being the detection of anomalies on Automatic Identification System (AIS) data from fishing ships to detect and/or anticipate abnormal fishing patterns, which is of key economic and environmental importance.