Comparative Analysis of NPU- Accelerated Edge Computing Platforms for Object Detection Model Implementation

Sammanfattning

As the demand grows for low-power, low-latency inference in applications such as surveillance, robotics, and embedded vision, selecting the appropriate combination of model and hardware becomes critical. However, few deployment-oriented studies provide a comprehensive evaluation of performance and long-term cost implications for running state-of-the-art object detection models on edge devices. This thesis explores the deployment and optimization of real-time object detection models on NPU-accelerated edge AI hardware platforms, focusing on the NXP i.MX 8M Plus with integrated NPU and Hailo-8L external NPU. The core objective of this work is to assess the inference performance, optimization effects, and total cost of ownership (TCO) for candidate edge AI platforms running YOLOv8 and YOLOv9 object detection models. The key challenge is to balance detection accuracy, runtime efficiency, and deployment cost within the limitations of edge computing hardware. To address this, the study benchmarks multiple configurations, including variations in input shape and camera stream resolution. The evaluation considers metrics such as frames per second (FPS), mean Average Precision (mAP), power consumption, and thermal behavior. Additionally, a five-year TCO analysis is conducted for a deployment scenario involving 100 devices, incorporating hardware pricing, expected failure rates, and energy costs. The findings indicate that Hailo-8L delivers higher inference throughput and better energy efficiency in terms of FPS per watt. However, the NXP platform proves more cost-effective over time due to its lower energy consumption and higher reported reliability. YOLOv8 models show strong post-quantization performance with minimal accuracy loss, making them well-suited for deployment on constrained devices. YOLOv9 models present varied outcomes, with YOLOv9m maintaining high accuracy after quantization, while YOLOv9s suffers significant degradation. This research provides a practical framework for selecting edge AI solutions based on specific trade-offs between performance, cost, and model robustness. It enables better decision-making for scalable deployment of intelligent vision systems and offers valuable benchmarks for future research in edge AI.