AERIS: An Adaptive, Battery-Aware Edge-AI Framework for Real-Time Survivor Detection in SAR Drones | RISE Research

This paper presents AERIS, an adaptive, battery-aware edge-AI framework for real-time survivor detection on search-and-rescue (SAR) drones, addressing the core constraint that a UAV's flight time and its onboard computation draw from a single limited battery. The authors trained a YOLOv8n object detector on a unified corpus of 24,437 aerial images from four public SAR datasets, then exported it into three deployment tiers of increasing computational cost: FP32 inference at 640px, INT8-quantized inference at 640px, and INT8-quantized inference at 320px. A physics-based controller projects remaining flight time at every frame using live propulsion and battery-state readings, favoring high-accuracy tiers while endurance allows and stepping down only as reserves become critical. Tested in a flight-to-depletion simulation against a static single-tier baseline, the adaptive controller reduced compute-energy consumption by 11.2% and processed nearly twice as many frames per flight, while INT8 quantization at full resolution matched full-precision accuracy to within 0.3 mAP points, showing quantization is a near-free efficiency gain compared to resolution reduction. Because propulsion dominates the total energy budget, tier-switching did not extend overall flight time; the contribution is optimizing energy consumption within a fixed flight duration rather than extending it. This efficiency gain carries a measurable accuracy cost: the adaptive strategy's time-weighted mAP50 (0.534) falls notably below the static baseline's (0.640), since nearly half of adaptive flight time is spent in the lowest-accuracy tier under the conservative safety margins used. The paper concludes that battery-state-aware inference scheduling is a viable, quantifiable design lever for resource-constrained SAR drones, and identifies the resulting accuracy-coverage tradeoff as the key open question for future mission-level evaluation.