ENHANCING THE EFFICIENCY OF VIDEO SURVEILLANCE SYSTEMS THROUGH A HYBRID METHOD OF KEY FRAME SELECTION AND DECISION INTERPRETATION

DOI: 10.31673/2412-4338.2025.038710

Abstract

Modern video surveillance systems generate significant volumes of data, complicating their effective real-time analysis. Existing automated anomaly detection methods are often computationally intensive and operate as "black boxes," limiting their practical application in critical areas such as public order maintenance and public safety. This article proposes a novel combined approach that addresses two key problems: inefficient processing of redundant data and insufficient transparency of artificial intelligence algorithms. The methodology is based on combining an innovative method for selecting informative frames with their subsequent processing by an interpretable detection model. The first stage involves optimizing feature selection using a hybrid algorithm that combines the InceptionV3 convolutional neural network with a genetic algorithm, enabling a 70-85% reduction in data volume while maintaining a 98% recall rate. The second stage ensures not only anomaly classification but also the generation of human-understandable explanations through the integration of explainable AI (XAI) methods, particularly Grad-CAM and guided backpropagation. Experimental validation on standard datasets demonstrates the advantages of the proposed approach compared to contemporary alternatives. The obtained results indicate a 3-5% improvement in classification accuracy while simultaneously reducing computational load. Furthermore, the system provides visual explanations of decisions in the form of heatmaps, enhancing trust in its operation. The proposed approach opens prospects for implementing efficient real-time video monitoring systems with decision justification capabilities.

Keywords: video surveillance, information systems, genetic algorithm, artificial intelligence, modeling, convolutional neural networks, video data processing, computer vision.