EUVIMEDEUVIMED entdecken
Crossref

Multi-View Edge Inference in Low-Altitude Airspace: Joint UAV Deployment and Model Architecture Optimization

Yao Tang, Guangxu Zhu, Chao Shen, Tse-Tin Chan, Tat-Ming Lok

IEEE Transactions on Mobile Computing · 2026-12 · 25 (12) · 22885-22899

DOI: 10.1109/tmc.2026.3725433 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Mobile Computing, 2026-12. Verfasst von Yao Tang und 4 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben. Gefördert von: Shenzhen Loop Area Institute, National Natural Science Foundation of China, Shenzhen Science and Technology Program.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The rapid emergence of the low-altitude economy (LAE) is enabling efficient and flexible applications via collaborative unmanned aerial vehicles (UAVs) operating in low-altitude airspace. In multi-view edge inference, UAVs capture data from diverse viewpoints and transmit features to an edge server, which fuses multi-view features for accurate recognition. However, the practical implementation is challenged by UAV-specific constraints and model architecture, notably UAV deployment and the view-pooling layer. In this paper, we investigate the joint design of UAV deployment and model architecture for multi-view edge inference in a target recognition task. To accommodate the UAV’s limited onboard resources, we adopt split inference. We first derive a closed-form accuracy bound that incorporates both the classification margin and the quantization error. We establish a closed-form relation that explicitly links UAV deployment and view-pooling layer placement to the classification margin. Then, we characterize the quantization error introduced by feature quantization during wireless transmission. Our analysis shows that jointly optimizing these factors yields more discriminative features and a more effective model architecture, thereby improving inference accuracy. Moreover, these design choices collectively shape inference latency, revealing a tradeoff between accuracy and latency. We formulate a joint UAV deployment and model architecture optimization problem that minimizes inference latency subject to an accuracy constraint. To solve the resulting mixed-integer non-convex problem, we develop an alternating optimization approach that decomposes it into tractable subproblems. The proposed split, view-pooling, quantization, and position (SVQP) scheme iteratively optimizes the corresponding decision variables. Simulation results demonstrate that the proposed SVQP scheme achieves the best tradeoff between inference accuracy and latency compared with baselines.

Förderung

  • Shenzhen Loop Area Institute
  • National Natural Science Foundation of China
  • Shenzhen Science and Technology Program
  • Guangdong Young Talent Research Project
  • Research Matching Grant Scheme from the Research Grants Council of Hong Kong

Zitieren

APA

Tang, Y., Zhu, G., Shen, C., Chan, T., & Lok, T. (2026). Multi-View Edge Inference in Low-Altitude Airspace: Joint UAV Deployment and Model Architecture Optimization. IEEE Transactions on Mobile Computing, 25(12), 22885-22899. https://doi.org/10.1109/tmc.2026.3725433

Vancouver

Tang Y, Zhu G, Shen C, Chan T, Lok T. Multi-View Edge Inference in Low-Altitude Airspace: Joint UAV Deployment and Model Architecture Optimization. IEEE Transactions on Mobile Computing. 2026;25(12):22885-22899. doi:10.1109/tmc.2026.3725433

BibTeX
@article{Tang2026,
  author = {Tang, Yao and Zhu, Guangxu and Shen, Chao and Chan, Tse-Tin and Lok, Tat-Ming},
  title = {Multi-View Edge Inference in Low-Altitude Airspace: Joint UAV Deployment and Model Architecture Optimization},
  journal = {IEEE Transactions on Mobile Computing},
  year = {2026},
  volume = {25},
  number = {12},
  pages = {22885--22899},
  doi = {10.1109/tmc.2026.3725433}
}
Kostenlos starten →