Towards Efficient Video Stream Analysis: A Distributed Deep Learning Framework: The DiVA Approach

Abstract

The advent of advanced computational devices and Neural Networks (NN) has triggered a paradigm shift in object detection, a key area of Artificial Intelligence (AI). This progress has significantly improved the accuracy of object identification in images, demonstrating the transformative power of deep learning. However, real-time video stream processing with deep learning models remains a challenge. This paper presents Distributed Video Analytics (DiVA), a scalable platform designed to address these issues using deep learning and event processing for real-time video analysis. It explores quantification techniques, optimization tools, and a high-level conceptual architecture to enhance video stream analysis. The study includes experiments evaluating the You Only Look Once version 8 small (YOLOv8s) model across various frameworks, hardware configurations, and optimization strategies. The results show substantial performance gains, particularly with Graphics Processing Unit (GPU) processing and advanced frameworks like NVIDIA Triton Server and Deepstream SDK, optimized with NVIDIA TensorRT and INT8 quantization. The findings highlight DiVA's effectiveness in improving performance, energy efficiency, and scalability for deep learning inference and model deployment. Notably, the best configuration achieved 47.2 frames per second (FPS), showcasing significant processing efficiency.