A Deep Learning based Integrated Memory Aware Twin AutoEncoder Network for Anomaly Detection in Video Surveillance on Edge Devices

Abstract

In an era where urban surveillance plays a crucial role in ensuring public safety, the rapid expansion of urban populations necessitates the advancement of surveillance technologies. The proliferation of resource-constrained Internet of Things (IoT) devices in recent years has posed significant challenges in managing efficient computation and real-time anomaly detection. In response to these challenges, this paper introduces the Adaptive Edge-Offload Anomaly Detection (AEAD) methodology, which offers a dynamic and adaptive approach to managing IoT device resources by making informed decisions regarding the offloading of computational tasks to edge servers. To detect anomalies, the Integrated Memory-Aware Twin Autoencoder Network (IMAN) is designed; IMAN comprises twin autoencoders that extract fused features based on appearance and motion, while a memory network is employed to select the most efficient features. By efficiently segmenting data and optimizing processing layers, AEAD enhances the accuracy of anomaly detection while minimizing energy consumption. The contributions of AEAD include its ability to strike a balance between local and edge processing based on real-time network conditions, ensuring that tasks are completed within predefined time constraints. Moreover, AEAD's adaptability empowers it to efficiently detect anomalies in scenarios such as video surveillance and sensor networks, making it a valuable asset for applications requiring enhanced security and surveillance capabilities. A comparative analysis of three datasets—University of California San Diego Pedestrian Dataset 2 (UCSD PED2), The Chinese University of Hong Kong Avenue Dataset (CUHK Avenue), and ShanghaiTech—reveals that the Proposed System (PS) methodology consistently outperforms the Existing System (ES) methodology. PS achieves Area Under the Curve (AUC) improvements of 7.16% on UCSD PED2, 11.306% on CUHK Avenue, and 6.760% on ShanghaiTech. These results underscore the superior effectiveness of PS in various anomaly detection scenarios.