Anomaly Detection in Traffic Applications: A Probabilistic Forecasting Approach Based on Object Tracking

Traffic anomaly detection is a critical challenge at the intersection of computer vision, machine learning, and intelligent transportation systems. This thesis addresses anomaly detection in static overhead traffic scenes, such as intersections and roundabouts, where anomalies appear often subtle at the pixel level due to high camera placement and dense traffic, and are often context-dependent and interaction-based. Unlike prior approaches based on video reconstruction or future-frame prediction, this thesis proposes a Probabilistic Movement Forecaster (PMF) that conditions on current detections and surrounding traffic agents to estimate the plausibility of near-future motion. The PMF outputs probabilistic forecasts of vehicle positions and derives anomaly scores from predicted and observed outcomes. Three innovations are introduced: (i) explicit context integration of neighboring agents, (ii) distance-aware regularization of covariance estimates, and (iii) asymmetric scoring to improve sensitivity to direction-specific uncertainties. These scores are aggregated into trajectory-level measures using ranking- based functions, with exponentially weighted averaging proving robust and effective across diverse scenes. To evaluate the approach, a dataset is created comprising three urban camera locations with over 1,000 annotated anomalies across 32 fine-grained categories and graded relevance levels. Evaluation with normalized discounted cumulative gain (NDCG) and area under the precision–recall curve (AU-PR) demonstrates that the PMF consistently detects and prioritizes critical anomalies, while remaining resilient to detection and tracking noise. The dataset, including complete background traffic, establishes a benchmark for future work that can be easily extended. This thesis shows that probabilistic motion forecasting, combined with severity-aware ranking metrics, provides a principled and practical basis for anomaly detection in complex traffic environments.