EMM-Det: Energy-Efficient Multidrone Tiny Object Detection by Memory-Enhanced Spiking Neural Networks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Unmanned Aerial Vehicles (UAVs) are crucial for safety and rescue but face limitations in payload, energy, and data privacy.
• Centralized training models for UAVs risk data leakage and privacy violations.

Purpose of the Study:

• To propose EMM-Det, a novel low-power, distributed object detection method for UAVs.
• To address energy constraints, enhance detection accuracy, and ensure data privacy in UAV operations.

Main Methods:

• Developed EMM-Det utilizing memory-enhanced spiking neurons with dynamic leakage constants for improved firing rates.
• Incorporated wavelet transform for multiscale feature encoding to enhance detection robustness.
• Leveraged crowdsourced perception and federated learning (FL) to boost data collection and mitigate privacy risks.

Main Results:

• EMM-Det achieved 81.8% mAP@50:95 detection accuracy on a constructed dataset.
• Demonstrated extremely low power consumption, outperforming other methods by 3.2% and traditional ANNs by 14.5%.
• Validated EMM-Det's balance of computational efficiency, noise resilience, and data privacy.

Conclusions:

• EMM-Det offers a promising solution for UAV object detection in environments with strict energy and privacy requirements.
• The method shows strong potential for real-world applications like community safety patrols and emergency rescue.