MDED-Framework: A Distributed Microservice Deep-Learning Framework for Object Detection in Edge Computing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Edge Computing

Background:

• Increasing data volumes and real-time processing needs drive demand for edge-based deep learning.
• Edge environments face resource constraints, requiring efficient deep learning model distribution.
• Distributing models demands careful resource allocation and lightweight designs to prevent performance loss.

Purpose of the Study:

• To propose a novel framework for easy deployment and distributed processing of deep learning models in edge computing.
• To address the challenges of resource limitations and performance degradation in edge AI.

Main Methods:

• Introduction of the Microservice Deep-learning Edge Detection (MDED) framework.
• Utilization of Docker containers and Kubernetes orchestration for deployment and distribution.
• Ensemble of high-level feature-specific networks (HFN) and low-level feature-specific networks (LFN) for pedestrian detection.

Main Results:

• Achieved a pedestrian detection speed of up to 19 FPS, meeting semi-real-time requirements.
• Demonstrated accuracy improvements up to AP50 and AP0.18 on the MOT20Det dataset.
• Successfully deployed a deep learning model in an edge computing environment.

Conclusions:

• The MDED framework facilitates efficient and accurate deep learning model deployment on edge devices.
• The proposed ensemble method enhances pedestrian detection performance in edge environments.
• MDED offers a viable solution for real-time AI applications at the edge.