Analysis of methods and technologies of intelligent video surveillance

Abstract

Intelligent video surveillance systems are one of the key areas of development of modern information technologies, as they provide automatic analysis of video streams, detection of objects and events in real time. Thanks to the development of deep learning methods and computer vision, it has become possible to create high-precision systems capable of operating in conditions of variable lighting, noise and complex background. At the same time, the task of increasing the efficiency of such systems by selecting optimal neural network architectures and hardware acceleration technologies that reduce frame processing latency and energy consumption on embedded devices (Edge AI) remains relevant. The purpose of the work is to analyze modern methods and technologies for creating intelligent video surveillance systems, determine the most effective architectures of deep neural networks for object detection and tracking tasks, and evaluate the possibilities of their optimization for real-time operation. The study considers convolutional neural network architectures (CNN) and transformer models (ViT, DETR), as well as hybrid approaches that combine spatial and temporal video analysis. A comparison of the YOLOv5/YOLOv8, OpenVINO, TensorRT, and RKNN Toolkit frameworks, which provide hardware acceleration on GPU and NPU platforms, is carried out. Additionally, the effectiveness of the DeepSORT and ByteTrack tracking algorithms, which provide stable tracking of objects in streaming video, was analyzed. The results of the study showed that the combination of compact CNN models with hardware-optimized libraries allows reducing the frame processing delay to 30 ms while maintaining detection accuracy of over 90%. The obtained conclusions confirm the feasibility of using hybrid architectures and hardware acceleration technologies to create effective intelligent video surveillance systems of the new generation