A Deep Learning-Based Face Mask Detector for Autonomous Nano-Drones (Student Abstract)

Eiman AlNuaimi, Elia Cereda, Rafail Psiakis, Suresh Sugumar, Alessandro Giusti, Daniele PalossiAAAI 2022 · DOI

Abstract

This paper presents a lightweight deep learning pipeline designed for real-time face mask detection on sub-10-cm nano-drones. Utilizing a fine-tuned and 8-bit quantized MobileNetV2 architecture, the system operates on ultra-low-power hardware with minimal accuracy loss. Experimental results demonstrate that the model achieves functional performance on-board the GAP8 SoC, enabling autonomous monitoring in indoor environments.

Introduction

The COVID-19 pandemic underscored the importance of face mask compliance in public spaces. While fixed surveillance systems are common, they lack the mobility required for comprehensive indoor monitoring. Nano-drones, such as the Crazyflie, offer a promising alternative due to their small size and ability to navigate safely near humans. However, deploying complex deep neural networks (DNNs) on these platforms is challenging due to severe resource constraints.

Overview of the nano-drone system and onboard face mask detection pipeline.

Methodology

The researchers employed the MobileNetV2 architecture, optimized for efficiency. The deployment pipeline involved three critical steps:

  1. Training Strategy: Comparing models trained from scratch against those fine-tuned from ImageNet weights.
  2. Quantization: Converting the 32-bit floating-point model to 8-bit integer precision using the NEMO library and PACT strategy to accommodate the GAP8 SoC's lack of floating-point hardware.
  3. On-board Deployment: Ensuring all computations occur locally to avoid latency and connectivity issues associated with off-board processing.

Dataset Development

A custom dataset was created to reflect the unique perspective of a nano-drone. It comprises approximately 19,000 images from 21 videos, categorized into three classes: Masked (M), Not Masked (NM), and No Person (NP). Data collection spanned various indoor environments to ensure model robustness.

Results and Discussion

The evaluation focused on the impact of transfer learning and quantization on model performance.

Qualitative results showing the model's classification scores across the three target classes compared to ground truth (GT).

Key findings include:

  • Transfer Learning is Essential: Fine-tuning was mandatory; the model trained from scratch achieved only 39% accuracy, whereas the fine-tuned version reached 64%.
  • Low Quantization Tax: The transition to 8-bit precision resulted in only a marginal drop in accuracy (from 64% to 61%) and AUC (from 0.81 to 0.80). This suggests that ultra-low-power hardware is capable of hosting complex AI.
  • Inference Speed: The quantized model is estimated to run at 3–4 frames per second on the GAP8 SoC, which is sufficient for low-speed autonomous flight and monitoring.

Conclusion

This work demonstrates the feasibility of running real-time face mask detection on ultra-low-power nano-drones. While the current 61% accuracy is a strong proof-of-concept, future work will focus on architectural optimizations for Parallel Ultra-Low Power (PULP) platforms and exploring more aggressive quantization techniques to further enhance efficiency and performance.

Summary updated

BibTeX
@inproceedings{alnuaimi2022deep,
  author = {Eiman AlNuaimi and Elia Cereda and Rafail Psiakis and Suresh Sugumar and Alessandro Giusti and Daniele Palossi},
  title = {{A Deep Learning-Based Face Mask Detector for Autonomous Nano-Drones (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21588},
}
Figure