Design and Evaluation of a Low‑Latency Drone‑Based Vision System

Abstract

This paper presents a controlled empirical study of face recognition under realistic UAV deployment constraints. We combine a live UAV streaming pipeline with a benchmark framework that fixes the detector, preprocessing, gallery construction, and similarity metric while varying only the recognition model. Seven pretrained recognizers, ranging from a 1.1M-parameter SFace model to a 65M-parameter ArcFace-R100 model, are evaluated under a simulated 400 MHz single-core CPU and 128 MB inference-memory budget on the DroneFace dataset, with VGGFace2 used as a generalization probe. We also report per-altitude, per-distance, and per-gender breakdowns. Three findings emerge. First, MobileFaceNet achieves 92.74% accuracy at 16.7 ms latency (59.7 FPS), within approximately three percentage points of the accuracy-best ArcFace-R50 model while running about 39× faster. Second, training-distribution alignment can be as important as architectural capacity: FaceNet performs best on VGGFace2 but degrades substantially on DroneFace, while ArcFace-based models show stronger performance under aerial conditions. Third, CPU-frequency scaling is non-uniform, with mid-weight models benefiting more clearly than very light models dominated by runtime overhead or very heavy models dominated by compute saturation. These results identify MobileFaceNet as the most practical deployment choice under the simulated UAV constraint, and live re-validation on 17 edge-case streaming clips preserves the Phase 2 ranking while exposing a sharp drop in absolute accuracy attributable to threshold-gated matching and to specific lighting and geometry cases.

Key findings

Limitations & future work

• The hardware budget is simulated with an x86 duty-cycle throttle, so latencies are good for ranking but not cycle-accurate for a real onboard processor. Next step: deploy MobileFaceNet on a physical ARM target.
• DroneFace has only 11 identities (7 male, 4 female), so per-gender gaps may reflect individual subjects rather than systematic bias, and it has no skin-tone, head-covering, or weather annotations.
• The live clip catalogue has no unknown-only clips, so false-alarm rate is not yet measured. A per-environment τ sweep is the natural next experiment.

Code

Citation

Copy@techreport{alghfeli2026dronevision,
  title       = {Design and Evaluation of a Low‑Latency Drone‑Based Vision System},
  author      = {Alghfeli, Alyaa and Almulla, Buthaina and Utemissova, Dilnaz
                 and Khanfar, Jenna and Yaqoob, Shamma},
  institution = {Mohamed bin Zayed University of Artificial Intelligence},
  address     = {Abu Dhabi, United Arab Emirates},
  year        = {2026}
}