Scalable clustering of segmented trajectories within a continuous time framework: application to maritime traffic data

Abstract

In the context of the surveillance of the maritime traffic, a major challenge is the automatic identification of traffic flows from a set of observed trajectories, in order to derive good management measures or to detect abnormal or illegal behaviours for example. In this paper, we propose a new modelling framework to cluster sequences of a large amount of trajectories recorded at potentially irregular frequencies. The model is specified within a continuous time framework, being robust to irregular sampling in records and accounting for possible heterogeneous movement patterns within a single trajectory. It partitions a trajectory into sub-trajectories, or movement modes, allowing a clustering of both individuals’ movement patterns and trajectories. The clustering is performed using non parametric Bayesian methods, namely the hierarchical Dirichlet process, and considers a stochastic variational inference to estimate the model’s parameters, hence providing a scalable method in an easy-to-distribute framework. Performance is assessed on both simulated data and on our motivational large trajectory dataset from the automatic identification system, used to monitor the world maritime traffic: the clusters represent significant, atomic motion-patterns, making the model informative for stakeholders.