Efficient Detection Framework Adaptation for Edge Computing: A Plug-and-play Neural Network Toolbox Enabling Edge Deployment

Jiaqi Wu, Shihao Zhang, Simin Chen, Lixu Wang, Zehua Wang, Wei Chen, Fangyuan He, Zijian Tian, F. Richard Yu, Victor C. M. Leung

Introduction

Deploying high-precision object detection models on resource-constrained edge devices—such as CCTV cameras or industrial micro-computers—presents a significant engineering challenge. Standard models are often too computationally expensive for real-time edge execution, while lightweight alternatives frequently sacrifice the accuracy required for critical safety tasks.

The Edge Detection Toolbox (ED-TOOLBOX) addresses this gap by providing a suite of plug-and-play neural network components designed to optimize popular frameworks like YOLO for edge environments. By introducing reparameterized backbones and sparse attention mechanisms, this framework achieves a 23% reduction in computational overhead (FLOPs) while simultaneously improving detection accuracy for complex industrial safety scenarios.

Edge Computing Architecture

Traditional cloud-based surveillance systems often suffer from high data transmission latency. The ED-TOOLBOX is designed for a decentralized architecture where processing occurs at the network's edge. This "Internet of Video Things" (IoVT) approach allows for real-time safety monitoring across various working scenes without the bottleneck of a centralized cloud control center.

System architecture showing the relationship between centralized cloud computing and localized edge computing nodes in a video surveillance system.

Methodology

The toolbox introduces three primary architectural innovations:

  1. Rep-DConvNet (Backbone): Utilizing structural reparameterization, this backbone employs a complex, multi-branch structure during the training phase to capture rich features. During inference, these branches collapse into a single, fast convolutional layer, maximizing deployment speed without losing learned capacity.
  2. Sparse Cross-Attention (SC-A) & Joint Module: To avoid the high computational cost of standard attention mechanisms, SC-A uses a localized-mapping approach. By performing pooling operations to create sparse feature representations, the model captures long-range dependencies across the network (Backbone, Neck, and Head) with minimal parameter overhead.
  3. Efficient Head: The standard dual-branch prediction head is replaced with a streamlined single-branch structure. Crucially, a specialized detection head for small-scale objects (160x160 feature maps) is integrated to detect fine details, such as helmet chin straps.

Experimental Results and the HBDD Dataset

A core contribution of this research is the Helmet Band Detection Dataset (HBDD). Unlike standard datasets that only verify helmet presence, HBDD focuses on a critical safety violation: unfastened chin straps. The dataset includes 8,100 images across four classes: no helmet, helmet worn correctly, helmet worn without strap tied, and worker's head.

Detection results from the ED-YOLO model across various industrial environments, demonstrating successful classification of helmet safety compliance.

In comparative testing, the ED-YOLO model (built using the toolbox) achieved 91.34% mAP on the HBDD dataset, significantly outperforming the YOLOv8-s baseline (82.15%). Furthermore, the model reduced parameter count by 17% and computational complexity by 23%, ensuring stable real-time performance in simulated IoVT environments.

Conclusion and Future Work

ED-TOOLBOX demonstrates that reparameterization and sparse attention are highly effective for edge-based safety monitoring. While currently optimized for Convolutional Neural Networks (CNNs), future iterations of the toolbox aim to support Transformer-based architectures and expand into tasks like instance segmentation and object tracking. The authors also intend to explore how these lightweight modules can facilitate the adaptation of large foundation models for specialized edge applications.

Summary updated

BibTeX
@article{wu2025efficient,
  author = {Jiaqi Wu and Shihao Zhang and Simin Chen and Lixu Wang and Zehua Wang and Wei Chen and Fangyuan He and Zijian Tian and F. Richard Yu and Victor C. M. Leung},
  title = {{Efficient Detection Framework Adaptation for Edge Computing: A Plug-and-play Neural Network Toolbox Enabling Edge Deployment}},
  journal = {IEEE Transactions on Mobile Computing},
  year = {2025},
  doi = {10.1109/tmc.2025.3612469},
}
Figure