AADC-Net: A Multimodal Deep Learning Framework for Automatic Anomaly Detection in Real-Time Surveillance

Automatic anomaly detection (AAD) has emerged as an advanced vision-based measurement method with diverse applications in healthcare and security. However, current AAD methods still face challenges related to data limitations and labeled-data imbalances, which limit the accuracy and reliability of AAD in real-life applications. Additionally, labeling and training large datasets for video anomaly detection (VAD) is computationally demanding and time-consuming. To address these challenges, this work introduces AADC-Net, a multimodal deep neural network for automated abnormal event detection and categorization. The key contributions of this research are as follows: 1) AADC-Net leverages pretrained large language models (LLMs) and vision-language models (VLMs) to mitigate VAD dataset limitations and imbalances; 2) a pretrained object detection model [DEtection TRansformer (DETR)] is integrated for visual feature extraction, eliminating the need for bounding box supervision; 3) the experimental results demonstrate the state-of-the-art (SOTA) performance of the proposed AADC-Net with an area under the curve (AUC) of 83.2% and an average precision (AP) of 83.8% on the public UCF-Crime and XD-Violence datasets, respectively; and 4) additionally, AADC-Net can be integrated into existing video surveillance systems, such as those in smart gyms and healthcare facilities, to automatically detect anomalies in real time with minimal supervision, enhancing security, monitoring, and reducing labor costs while minimizing human error. In summary, our results demonstrate that AADC-Net not only achieves high accuracy in anomaly detection but also provides a practical solution for real-world surveillance applications.