EFFICIENT EDGE AI FRAMEWORKFOR REAL-TIME IOT VISION SYSTEM USING DEEP LEARNING

Abstract

Modern vision systems are widely used in autonomous vehicles, smart surveillance, healthcare imaging, industrial monitoring, and IoT-based applications. These systems continuously process large amounts of visual data and require fast and reliable decision making. Traditional cloud-based processing often faces problems such as high latency, bandwidth usage, privacy risks, and unstable network connectivity. Edge Artificial Intelligence (Edge AI) has emerged as an effective solution by performing inference closer to the data source. This work studies efficient deployment of deep learning based vision models on resource-constrained edge devices. The work focuses on understanding how different optimisation techniques affect latency, throughput, memory usage, and model size during deployment. A comparative framework is proposed for four major AI-IoT architectures: cloud-only, edge-only, edge-cloud hybrid, and federated learning. The study also reviews three major model compression approaches: pruning, quantisation, and knowledge distillation. An experimental analysis is carried out using the YOLOv8n object detection model in three configurations: PyTorch FP32, ONNX FP32, and ONNX INT8. The experiments are performed on a CPU-only environment using a subset of the Intel IoT DevKit traffic dataset. Results show that INT8 quantisation This has a significant impact on the size of the model, but has no impact on the quality of the inference. The CPU hardware used for testing. The results are evidence that The advantages of model compression are very much dependent on the The underlying hardware and runtime environment. The study finds that for effective edge-AI deployment, four key components must be in place: careful balance between accuracy, latency, hardware capability, and resource efficiency.