Maritime Vessel Tracking

The Automatic Identification System (AIS) provides time stamped vessel positions and kinematic reports that enable maritime authorities to monitor traffic. We consider the problem of relabeling AIS trajectories when vessel identifiers are missing, focusing on a challenging nationwide setting in which tracks are heavily downsampled and span diverse operating environments across continental U.S. waters. We propose a hybrid pipeline that first applies a physics-based screening step to project active track endpoints forward in time and select a small set of plausible ancestors for each new observation. A supervised neural classifier then chooses among these candidates, or initiates a new track, using engineered space time and kinematic consistency features. On held out data, this approach improves posit accuracy relative to unsupervised baselines, demonstrating that combining simple motion models with learned disambiguation can scale vessel relabeling to heterogeneous, high volume AIS streams.

💡 Research Summary

The paper tackles the problem of relabeling Automatic Identification System (AIS) trajectories when vessel identifiers (MMSI) are missing, a task essential for maritime traffic monitoring and threat detection. While prior work has relied on unsupervised clustering or simple physics‑based heuristics, those approaches struggle on a nationwide scale where data are heavily down‑sampled (approximately one report per vessel every 30 minutes) and the operating environment is highly heterogeneous (open ocean, coastal waters, rivers, ports, and docks).

To address this, the authors propose a hybrid pipeline that combines a physics‑based candidate screening step with a supervised neural classifier. The pipeline works as follows: for each new AIS observation (a “posit”) the system projects all active track endpoints forward in time using continuous‑time kinematic models (constant velocity, constant turn‑rate, and constant acceleration). The projected state is compared to the observation in a local along‑track / cross‑track frame, yielding an anisotropic Gaussian error model with separate variances along and across the heading. Mahalanobis distance and an angular consistency gate (≤ 85°) are applied to prune implausible links. A fast approximate MAP score that also incorporates turn‑rate, speed‑difference, and prior endpoint density is computed, and the top k = 16 candidates plus a “New Vessel” hypothesis are retained.

Each candidate is then described by a 231‑dimensional feature vector that captures (i) raw residuals (e∥, e⊥), time gap, required speed, and turn‑rate; (ii) forward and backward projection errors and their symmetrized averages; (iii) heading and speed consistency, jitter indicators; and (iv) contextual information such as endpoint density and time‑of‑day. A fully‑connected neural network with three hidden layers (sizes 2000, 2000, 1000) processes these vectors and outputs logits for the k candidates and the New Vessel class. The model is trained with cross‑entropy loss, label smoothing (ε = 0.05), mixed‑precision Adam optimizer, and temperature scaling for calibrated probabilities at inference.

The authors evaluate the method on a publicly available AIS dataset from Marine Cadastre covering 2,482 days of U.S. waters. After down‑sampling, adding modest spatio‑temporal noise, and removing stopped‑vessel reports, the dataset contains only active trajectories. The evaluation metric is “posit accuracy”: for each AIS point, one point is awarded for correctly identifying its predecessor and one for its successor; the total is normalized by the number of points.

Baseline methods include the original unsupervised CBTR (physics‑aware greedy linking) achieving 0.33 posit accuracy, the A‑TD 2025 challenge baseline (0.44), and several standard multi‑target tracking approaches (Kalman filter + nearest‑neighbor, DBSCAN, hierarchical clustering). The proposed hybrid system reaches a mean classification accuracy of 0.7452 (F1 ≈ 0.73) on a held‑out test set and improves overall posit accuracy to 0.53, a substantial gain over the baselines. An oracle that always selects the true candidate among the screened set would achieve 0.85 posit accuracy, indicating that both screening and classification contribute to the remaining error gap.

Performance varies by geography: open‑water regions exhibit the highest accuracy, while dense port, river, and dock areas suffer more mis‑links due to many plausible ancestors in close spatio‑temporal proximity. Error analysis shows that in high‑density zones the physics‑based screen sometimes discards the correct ancestor, and the neural classifier can struggle to disambiguate among the remaining candidates.

The discussion highlights two bottlenecks: (1) candidate screening may be too aggressive in congested waterways, and (2) the feature set lacks explicit waterway geometry and port‑specific priors. Future work is suggested to incorporate navigable polygon constraints, richer vessel behavior models (stop‑and‑go, lane‑following), and adaptive screening that expands the candidate set when uncertainty is high.

In conclusion, the study demonstrates that a simple physics‑based pruning step followed by a data‑driven neural disambiguation can scale vessel relabeling to nationwide, heavily down‑sampled AIS streams, achieving a notable 0.53 posit accuracy. While still modest compared to the near‑perfect results on localized datasets, the approach provides a solid foundation for further improvements in high‑traffic maritime environments. All code, preprocessing scripts, and the training dataset are publicly released.