Dual-Modality Deep Feature-based Anomaly Detection for Video Surveillance

Detecting anomalies in videos is not only crucial but also an intriguing task in surveillance systems. It is a sequential modeling problem in nature that requires careful selection of spatial and temporal dependent patterns from a sequence of frames. There are several research works from traditional approaches to modern deep learning-based techniques introduced to address this problem. However, there is a huge demand for research and development to ameliorate the performance of the existing solutions. In response to that, this study proposes an improved video anomaly detection model using deep features extracted from a dual-modality input representation. The proposed model demonstrates effectiveness in the benchmark–UCF crime dataset by achieving the best AUC of 87.52%, which is ≈ 12.3% improvement compared to a baseline. The application aspect of this work includes strengthening the security measures in common places, viz. airports, banks, public transits, schools, and shopping complexes by detecting aberrational or suspicious activities in surveillance videos.