Network-Aware 5G Edge Computing for Object Detection: Augmenting Wearables to "See" More, Farther and Faster

Zhongzheng Yuan, Tommy Azzino, Yu Hao, Yixuan Lyu, Haoyang Pei, Alain Boldini, Marco Mezzavilla, Mahya Beheshti, Maurizio Porfiri, Todd E. Hudson, William H. Seiple, Yi Fang, Sundeep Rangan, Yao Wang, John-Ross Rizzo

Introduction

Smart wearables designed for the Blind and Visually Impaired (BVI) face a critical "power vs. performance" trade-off. To navigate complex urban environments safely, a user needs a system that can detect small obstacles at a distance and provide a wide, ideally 360-degree, field of view. However, current systems like the VIS⁴ION instrumented backpack rely on local embedded processors (such as the NVIDIA Jetson), which are severely limited by battery life and thermal constraints. These devices are typically restricted to processing low-resolution video (WVGA) at low frame rates (~10 FPS), resulting in "tunnel vision" and a dangerously short detection range.

This paper demonstrates how 5G millimeter-wave (mmWave) and mobile edge computing (MEC) can revolutionize these assistive technologies by offloading heavy vision tasks to the network. By moving computation away from the device, the authors achieve high-resolution, multi-camera object detection with sub-100ms latency—capabilities that are currently impossible with local hardware.

System Architecture and Methods

The researchers proposed a full-stack system that leverages the massive uplink bandwidth of 5G to stream high-definition video to an edge server. The core of this approach is an Adaptive Offloading scheme that dynamically adjusts the number of camera feeds and their resolution based on real-time network monitoring.

The system architecture illustrates the flow from wearable camera feeds to local or edge processing, mediated by network monitoring.

To evaluate this system, the authors employed a rigorous simulation framework:

  • Vision Analysis: They evaluated the YOLOv5s object detection model across various spatial resolutions (from WVGA to 2.2K) to find the optimal balance for navigation safety.
  • Wireless Simulation: Using the ns-3 network simulator, they modeled a "Last Mile" pedestrian commute in a dense urban environment (Manhattan).
  • Ray-Tracing: To ensure high realism, they used Remcom Wireless InSite for ray-tracing, combined with detailed 3D city models that include building materials and vegetation.

The NYU-NYC StreetScene Dataset

A significant contribution of this work is the introduction of the NYU-NYC StreetScene dataset. Unlike standard computer vision datasets, this collection is specifically tailored for BVI navigation. It features high-resolution videos captured from a pedestrian's perspective in New York City, with manual labels for 15 object classes such as traffic lights, pedestrians, and fire hydrants. Crucially, the dataset includes depth estimation to evaluate how video resolution affects the "reliable detection range."

A pedestrian using the wearable system alongside a typical urban scene from the NYU-NYC StreetScene dataset.

Key Insights: Why 5G mmWave?

The study highlights several critical findings regarding the necessity of high-frequency 5G:

  1. Resolution Equals Range: Increasing resolution from WVGA to 1080p doubles the reliable detection range for pedestrians from approximately 6 meters to 12 meters. For a BVI user, those extra six meters represent vital seconds of reaction time.
  2. The 100ms Physiological Limit: The authors argue that total end-to-end latency must stay under 100ms. This is based on the duration of a human "saccade" (fast eye movement); anything slower feels disconnected from reality and is ineffective for real-time navigation.
  3. mmWave is a Necessity: While 4G LTE and sub-6-GHz 5G are more reliable, they lack the uplink capacity to support even a single 1080p stream at 30 FPS. 5G mmWave is the only technology capable of the >100 Mbps speeds required for an omnidirectional view.

Results and Performance

The results of the edge-offloading approach are compelling. By utilizing an edge server with an RTX 8000 GPU, the system achieved 1080p processing at over 30 FPS, whereas local processing on a Jetson Xavier NX was limited to WVGA at roughly 13 FPS.

In terms of latency, the 5G mmWave + LTE configuration achieved a median round-trip time of 15ms, contributing to a total system delay of 84ms—successfully meeting the 100ms target. While mmWave is susceptible to blockage, the adaptive scheme allowed the system to maintain "good" performance nearly 100% of the time by dropping resolution or camera count during poor signal periods.

Limitations and Future Work

Despite the performance gains, challenges remain. The 5G mmWave radio itself is power-hungry, and a full power-profile analysis is required to ensure long-term wearable viability. Future work will also explore "cooperative sensing," where data from other users or smart city infrastructure helps mitigate the effects of physical signal blockage. Additionally, the authors suggest that training models specifically on compressed, low-bitrate video could further improve accuracy without increasing bandwidth requirements.

Summary updated

BibTeX
@article{yuan2022network,
  author = {Zhongzheng Yuan and Tommy Azzino and Yu Hao and Yixuan Lyu and Haoyang Pei and Alain Boldini and Marco Mezzavilla and Mahya Beheshti and Maurizio Porfiri and Todd E. Hudson and William H. Seiple and Yi Fang and Sundeep Rangan and Yao Wang and John-Ross Rizzo},
  title = {{Network-Aware 5G Edge Computing for Object Detection: Augmenting Wearables to "See" More, Farther and Faster}},
  journal = {IEEE Access},
  year = {2022},
  doi = {10.1109/access.2022.3157876},
}
Figure