A GELU-ACTIVATED VIDEO ANOMALY DETECTION MODEL ON EDGE-AI DEVICES

Kahlil Muchtar^1,2, Al Bahri¹, Yudha Nurdin¹, Oky Firmansyah³, Alvin Prayuda Juniarta Dwiyantoro⁴, Martha Arbayani Zaidan^5,6, Abdallah Namoun^7,8, and Chih-Yang Lin⁹

¹ Department of Electrical and Computer Engineering, Universitas Syiah Kuala, Aceh, Indonesia

² Data Science and Artificial Intelligence Research Center, Universitas Syiah Kuala, Aceh, Indonesia

³ Nodeflux, Indonesia

⁴ Google, Indonesia

⁵ Department of Computer Science, Faculty of Science, University of Helsinki, PL 64, FI-00014, UHEL, Helsinki, Finland

⁶ Institute for Atmospheric and Earth System Research (INAR/Physics), Faculty of Science,

University of Helsinki, PL 64, FI-00014, UHEL, Helsinki, Finland

⁷ AI Center, Islamic University of Madinah, Madinah 42351, Saudi Arabia

⁸ Faculty of Computer and Information Systems, Islamic University of Madinah, Madinah 42351, Saudi Arabia

The growing implementation of Intelligent Surveillance Systems (ISS) has heightened the need for Video Anomaly Detection (VAD) to enhance security, reduce labor costs, and improve energy efficiency. Previous techniques typically demand substantial computational power and are limited to offline operations, rendering them unsuitable for online or real-time VAD surveillance applications. This paper introduces a VAD system based on Edge AI, which integrates a lightweight separable 3D-based Convolutional Neural Network (S3D CNN) extractor utilizing the Gaussian Error Linear Unit (GELU) activation function, resulting in an enhanced VAD system. We conduct a comprehensive analysis of the performance of our proposed solution on two well-known edge devices: the Jetson Orin Nano and the ASUS NUC Performance. In terms of dataset evaluations, two benchmark datasets (UCF-Crime and XD-Violence) are assessed, demonstrating that our proposed work can attain 86% and 91% of AUC (Area Under the Curve), respectively. Supplementary materials are available at https://edgevad.comvislab-usk.org/.