Date of Graduation


Document Type


Degree Type



Statler College of Engineering and Mineral Resources


Industrial and Managements Systems Engineering

Committee Chair

Imtiaz Ahmed

Committee Member

Zeyu Liu

Committee Member

Alan McKendall


In marine surveillance, a crucial task is distinguishing between normal and abnormal vessel movements to timely identify potential threats. Subsequently, the vessels need to be monitored and tracked until necessary action can be taken. To achieve this, a track association problem is formulated where multiple vessels' unlabeled geographic and motion parameters are associated with their true labels. These parameters are typically obtained from the Automatic Identification System (AIS) database, which enables real-time tracking of marine vessels equipped with AIS. The parameters are time-stamped and collected over a long period, and therefore, modeling the inherent temporal patterns in the data is crucial for successful track association. The problem is further complicated by infrequent data collection (time gap) and track overlaps.

Traditionally, physics-based models and Kalman-filtering algorithms are used for tracking problems. However, the performance of Kalman filtering is limited in the presence of time-gap and overlapping tracks, while physics-based models are unable to model temporal patterns. To address these limitations, this work employs LSTM, a special neural network architecture, for marine vessel track association. LSTM is capable of modeling long-term temporal patterns and associating a data point with its true track. The performance of LSTM is investigated, and its strengths and limitations are identified. To further improve the performance of LSTM, an integration of the physics-based model and LSTM is proposed. The performance of the joint model is evaluated on multiple AIS datasets with varying characteristics.

According to the findings, the physics-based model performs better when there is very little or no time gap in the dataset. However, when there are time gaps and multiple overlapping tracks, LSTM outperforms the physics-based model. Additionally, LSTM is more effective with larger datasets as it can learn the historical patterns of the features. Nevertheless, the joint model consistently outperforms the individual models by leveraging the strengths of both approaches. Given that the AIS dataset commonly provides a long stretch of historical information with frequent time gaps, the combined model should improve the accuracy of vessel tracking.