Pyramidal temporal frame prediction for efficient anomalous event detection in smart surveillance systems

Abstract

Video anomaly detection plays a pivotal role within contemporary intelligent surveillance systems, holding significant utility in the identification and mitigation of anomalies within smart environments. The task of anomaly detection is made difficult by the presence of complex environments and varying perspectives, even in diverse situations. Furthermore, the lack of labeled data in this field presents additional challenges. In this article, we present a robust pyramidal frame prediction architecture coupled with an efficient anomaly detection mechanism. The foundation of the frame prediction module rests upon a combination of localized predictors, incorporating bidirectional sampling techniques and a proficient discriminator-sharing mechanism. For the purpose of anomaly detection, an innovative loss function is employed, tailored to accommodate multi-instance learning scenarios. Departing from the conventional treatment of anomalies as a binary classification problem, we adopt a regression-oriented approach, thereby enabling a nuanced understanding of disruptive occurrences at a granular level, whilst facilitating an assessment of the severity associated with anomalous events. The amalgamation of these dual modules culminates in a distinctive and robust framework adept at comprehending anomalies within intricate scenarios. Empirical assessments conducted across benchmark datasets within this domain serve to underscore the efficacy of the proposed architectural paradigm across a spectrum of environments.