A Comprehensive Review of Multimodal Weakly Supervised Video Anomaly Detection

Video Anomaly Detection (VAD) has emerged as a critical research area in intelligent surveillance, public safety, industrial monitoring, and smart city applications. Traditional supervised approaches rely on frame-level annotations, which are labor-intensive, time-consuming, and expensive to obtain, thereby limiting their scalability in large-scale surveillance systems. Weakly Supervised Video Anomaly Detection (WSVAD) addresses this challenge by utilizing only video-level labels while enabling effective temporal anomaly localization. In recent years, multimodal learning has significantly improved detection performance by integrating complementary information from visual, audio, and semantic modalities. This paper presents a comprehensive review of multimodal WSVAD techniques, emphasizing recent advancements in Multiple Instance Learning (MIL), vision-language models, audio-visual learning, knowledge transfer, and semantic alignment strategies. Representative methods are systematically analyzed and compared with respect to their methodologies, benchmark datasets, evaluation metrics, strengths, and limitations. Furthermore, major research challenges, including modality imbalance, limited multimodal datasets, computational complexity, and domain generalization, are discussed. Finally, emerging research directions involving foundation models, multimodal large language models (MLLMs), self-supervised learning, federated learning, and explainable artificial intelligence (XAI) are highlighted to support the development of scalable, robust, and interpretable anomaly detection systems.