Hybrid AI framework for efficient anomaly detection in video surveillance data.

Bondarchuk, Andrii and Bushma, Oleksandr and Dovzhenko, Tymur and Hashko, Andrii (2025) Hybrid AI framework for efficient anomaly detection in video surveillance data. ScientificWorldJournal, 1 (33). pp. 188-195. ISSN 2663-5712

Text 189-195-PB.pdf Download (2MB)

Abstract

Contemporary video surveillance infrastructure produces substantial data streams,posing challenges for efficient real-time processing. Current automated anomaly detection techniques frequently demand extensive computational resources and function as opaque "black box" systems, constraining their deployment in critical domains including public security and safety monitoring. This work introduces an integrated methodology tackling two fundamental limitations: ineffective handling of superfluous visual data and lack of algorithmic transparency in artificial intelligence systems. The proposed framework merges an advanced informative frame selection technique with interpretable detection model processing. The initial phase employs a hybrid optimization approach integrating InceptionV3 convolutional neural networks with genetic algorithms, achieving 70-85% data reduction while preserving 98% recall performance. The subsequent phase delivers not only anomaly classification but also produces comprehensible decision explanations via explainable AI (XAI) integration, specifically utilizing Grad-CAM and guided backpropagation techniques. Experimental evaluation on benchmark datasets confirms the superiority of the proposed method over contemporary solutions. Results demonstrate 3-5% enhancement in classification precision coupled with reduced computational requirements. Additionally, the system generates visual decision rationalizations through heatmap representations, thereby increasing operational trustworthiness. This integrated framework facilitates the deployment of effective real-time video analysis systems that provide comprehensive decision transparency and operational accountability.

Item Type:	Article
Uncontrolled Keywords:	artificial intelligence; video surveillance; information systems; genetic algorithm; modeling; computer vision; video surveillance data
Subjects:	Статті у періодичних виданнях > Інші (не входять ні до фахових, ні до баз даних)
Divisions:	Факультет інформаційних технологій та математики > Кафедра комп'ютерних наук
Depositing User:	професор Андрій Бондарчук
Date Deposited:	07 Jan 2026 09:18
Last Modified:	20 Apr 2026 07:01
URI:	https://elibrary.kubg.edu.ua/id/eprint/56071

Actions (login required)

View Item