Abstract
Anomaly detection in surveillance footage is crucial for ensuring protection and safety standards, as it enables the timely identification of unusual or suspicious activities. Recent literature has shown the emergence of graph and hypergraph (HG)-based matching algorithms for object tracking in video frames, facilitating anomaly detection. While these techniques incorporate sampling methods to enhance pace, the task of balancing accuracy and efficiency remains unresolved, particularly when detecting anomalies during simultaneous object tracking. This paper addresses this gap by proposing a unified framework that integrates Game-Theoretic Hypergraph Matching (GTHG) with a Convolutional Autoencoder (CAE). Unlike existing methods that treat tracking and detection separately, the proposed approach combines structural consistency and appearance reconstruction to improve both detection accuracy and computational performance. The proposed method has been tested with a chain of benchmarked films and video clips, and a detailed account of matching between successive frames has been provided. Evaluation metrics, including the Regularity Score, Receiver Operating Characteristics (ROC), and Area Under the Curve (AUC), assess the accuracy of anomaly detection across multiple datasets. Our method achieves an area under the curve (AUC) of 88.7%, 91.2%, and 86.6% on the UCSD Ped1, UCSD Ped2, and CUHK Avenue Datasets, respectively, surpassing the performance of many existing models.