A Vehicle-Edge-Cloud Framework for Computational Analysis of a Fine-Tuned Deep Learning Model

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Cooperative, Connected, and Automated Mobility (CCAM) infrastructure is crucial for autonomous vehicle (AV) environmental perception in urban areas.
• Efficient deployment of machine learning (ML) and deep learning (DL) models is essential for enhancing AV performance within CCAM frameworks.
• The selection of computational processing layers and specific DL models impacts AV capabilities in complex environments.

Purpose of the Study:

• To propose a computational framework for deploying DL models in CCAM infrastructure.
• To analyze the effectiveness of a custom-trained YOLOv8 DL model across diverse devices and layers in a vehicle-edge-cloud architecture.
• To investigate the trade-offs between DL model accuracy and execution time during deployment within the layered framework.

Main Methods:

• Developed a computational framework for deploying DL models within a vehicle-edge-cloud layered architecture.
• Evaluated a custom-trained YOLOv8 DL model across various computational layers and devices.
• Measured performance metrics including accuracy (mAP@0.5) and inference time under different environmental conditions and hardware configurations.

Main Results:

• DL model performance metrics (e.g., 0.842 mAP@0.5) remained consistent across different device types within any layer of the framework.
• Inference times for object detection tasks decreased under varied environmental conditions.
• Specific hardware configurations showed significant reductions in inference time: Jetson AGX (non-GPU) reduced time by 72% compared to Raspberry Pi (non-GPU), and Jetson AGX Xavier (GPU) reduced time by 90% compared to Jetson AGX ARMv8 (non-GPU).

Conclusions:

• The choice of computational layer and device significantly impacts DL model inference time, not its accuracy, in CCAM systems.
• Researchers and practitioners can leverage these findings to optimize DL model deployment for production AV systems.
• Selecting appropriate hardware and deployment environments is key to balancing performance and efficiency in AVs.