Sparsity-driven Anomaly Detection for Ship Detection and Tracking in Maritime Video

摘要

This work examines joint anomaly detection and dictionary learning approaches for identifying anomalies in persistent surveillance applications that require data compression. We have developed a sparsity-driven anomaly detector that can be used for learning dictionaries to address these challenges. In our approach, each training datum is modeled as a sparse linear combination of dictionary atoms in the presence of noise. The noise term is modeled as additive Gaussian noise and a deterministic term models the anomalies. However, no model for the statistical distribution of the anomalies is made. An estimator is postulated for a dictionary that exploits the fact that since anomalies by definition are rare, only a few anomalies will be present when considering the entire dataset. Prom this vantage point, we endow the deterministic noise term (anomaly-related) with a group-sparsity property. A robust dictionary learning problem is postulated where a group-lasso penalty is used to encourage most anomaly-related noise components to be zero. The proposed estimator achieves robustness by both identifying the anomalies and removing their effect from the dictionary estimate. Our approach is applied to the problem of ship detection and tracking from full-motion video with promising results.