Drone-First-Aid SAD Systems

• Drone-First-Aid SAD is a system integrating multi-sensor drones with AI-based perception for rapid emergency assessment and precise aid delivery.
• It employs heterogeneous UAV architectures, high-resolution sensing, and deep neural networks for road, building, and person detection with real-time processing.
• Optimized network planning and layered safety protocols ensure coordinated drone operations, resilient response, and secure autonomous payload delivery.

Drone-First-Aid SAD (Search-and-Deliver) refers to integrated systems in which autonomous or semi-autonomous unmanned aerial vehicles (UAVs, commonly termed drones) conduct rapid search, assessment, and targeted aid delivery in emergency or disaster contexts. SAD systems fuse high-throughput perception, onboard real-time inferencing (often deep learning–based), communication, and precision delivery payloads, enabling faster, safer, and more scalable deployment of critical interventions before or in parallel with human response.

1. System Architectures: Platforms, Sensing, and Compute

State-of-the-art SAD system architectures employ heterogeneous drone fleets optimized for both rapid area surveying and targeted delivery. Typical platforms include fixed-wing UAVs for wide-area orthomosaic imaging (e.g., MACS-Micro camera at ≈80 km/h, 200 m AGL) and VTOL multirotors or helicopters (e.g., superARTIS) for vertical search and box-drop payload deployment (Merkle et al., 2023).

Onboard sensor suites are tailored to mission requirements:

• RGB and multispectral cameras (3 cm GSD) for high-resolution visual data.
• GNSS+IMU for accurate geo-referencing.
• Thermal/IR imagers and gas/LiDAR sensors when human detection in low-visibility or hazardous environments is critical (Mnaouer et al., 2021).

Processing is typically split between:

• Embedded GPUs (e.g., Jetson AGX Xavier, Jetson Nano) for onboard, float16-optimized CNN inference (real-time person/fire detection).
• Field-deployed high-end laptops or edge clusters (e.g., RTX 2080 Super) for near-real-time mosaicking and segmentation.
• Protocol Buffers and internal CAN/IPC minimize latency in data interchange.

Multi-drone coordination, coverage, and redundancy are supported by modular communication modules, including COTS radio, LoRaWAN (0.3–50 kbps, ≈15 km), and WiFi/4G LTE (5+ Mbps, <200 ms latency) (Mnaouer et al., 2021).

2. Remote-Sensing, Perception, and Data Processing Pipelines

Image acquisition in SAD leverages high-throughput frame rates (≥2 Hz), geo-tagging, and quantile-based normalization/downsizing to control outlier distributions and harmonize super-resolution imagery with deep model training regimes (Merkle et al., 2023).

Situation assessment relies on a three-headed CNN pipeline:

• Road Segmentation: Dense-U-Net-121, optimized with pixel-wise cross-entropy loss,

$$\mathcal{L}_{road}(\theta) = -\sum_{i\in\Omega} \bigl[y_i\log \hat y_i+(1-y_i)\log(1-\hat y_i)\bigr].$$

• Building Segmentation: HRNet (4 multi-resolution streams), trained with OHEM-enhanced cross-entropy,

$$\mathcal{L}_{bldg}(\theta) = \frac{1}{|\mathcal{B}_{hard}|} \sum_{i\in \mathcal{B}_{hard}} -[y_i\log \hat y_i + (1-y_i)\log(1-\hat y_i)].$$

• Person Detection: Modified YOLOv3 with anchor adaptation, optimized jointly for objectness and CIoU bounding-box regression,

$$\mathcal{L}_{person}(\theta) = \sum_{j}\ell_{obj}(p_j, \hat p_j) + \lambda\sum_{j}[1 - \mathrm{CIoU}(b_j,\hat b_j)].$$

The combined training objective is:

$$\min_{\theta}\;\mathcal{L}(\theta),\quad \mathcal{L} = \mathcal{L}_{road} + \mathcal{L}_{bldg} + \mathcal{L}_{person}.$$

Training employs augmentation (flip, rotate, rescale, brightness jitter) and architecture-specific optimizers (ADAM, SGD+Nesterov), with careful tuning of learning rates and weight decay.

Onboard pipelines operate on 416×416 px tiles (≈3 cm GSD), yielding inference latencies of ≈2 s per 16 MP image with negligible drop in detection AP on embedded hardware (Merkle et al., 2023). Coarse-to-fine image fusion overlays segmentation/detection outputs as GIS-ready GeoTIFFs, supporting both automated rerouting and human-in-the-loop review.

3. Autonomous Aid Delivery: Logic, Navigation, and Control

Payload delivery deploys strict safety and autonomy logic. Geo-referenced bounding-boxes with confidence scores define dynamic drop-zone clearance:

• For a circular drop zone $D$ of radius $R$ centered at $p^*$, clear if $\max_{c_k\in D} s_k < \tau$ (no bounding box above threshold), else abort and hold pattern.
• Guidance is based on minimum-snap polynomial trajectory optimization for waypoints:

$$\min_{q(t)} \int_{t_0}^{t_f} \left|\frac{d^4q}{dt^4}(t)\right|^2 dt, \text{ s.t.}~ q(t_0)=q_0,~ q(t_f)=q_f,~...$$

• Low-level control realizes PID laws for position hold:

$$\mathbf{u}(t) = K_p\,\mathbf{e}(t) + K_i \int_{0}^{t}\mathbf{e}(\tau)d\tau + K_d\,\dot{\mathbf{e}}(t),~~ \mathbf{e} = \mathbf{p}_{cmd}-\mathbf{p}_{est}.$$

• Payload actuation ensures feedback on door/hatch state for mission logging.

Fire-oriented SAD variants extend delivery logic for fire extinguishing: drones deploy suppression “balls” via AI-mapped fire-density targeting, alongside first-aid kits, coordinated with in-building sensor triggers and dense real-time mapping (Mnaouer et al., 2021).

4. Network Optimization, Coverage, and Response Time Modeling

Network-level SAD system design exploits mathematical programming (p-median or integer-linear models) to optimize drone base locations, fleet sizing, and assignment, subject to response-time thresholds and real-world constraints.

For urban/suburban scenarios, a location-queuing formulation integrates travel times ($$\mathcal{L}_{bldg}(\theta) = \frac{1}{|\mathcal{B}_{hard}|} \sum_{i\in \mathcal{B}_{hard}} -[y_i\log \hat y_i + (1-y_i)\log(1-\hat y_i)].$$0), base availability ($$\mathcal{L}_{bldg}(\theta) = \frac{1}{|\mathcal{B}_{hard}|} \sum_{i\in \mathcal{B}_{hard}} -[y_i\log \hat y_i + (1-y_i)\log(1-\hat y_i)].$$1), and baseline EMS times ($$\mathcal{L}_{bldg}(\theta) = \frac{1}{|\mathcal{B}_{hard}|} \sum_{i\in \mathcal{B}_{hard}} -[y_i\log \hat y_i + (1-y_i)\log(1-\hat y_i)].$$2), minimizing total drones while achieving specified response-time gains:

$$\mathcal{L}_{bldg}(\theta) = \frac{1}{|\mathcal{B}_{hard}|} \sum_{i\in \mathcal{B}_{hard}} -[y_i\log \hat y_i + (1-y_i)\log(1-\hat y_i)].$$3

$$\mathcal{L}_{bldg}(\theta) = \frac{1}{|\mathcal{B}_{hard}|} \sum_{i\in \mathcal{B}_{hard}} -[y_i\log \hat y_i + (1-y_i)\log(1-\hat y_i)].$$4 encodes improvement over baseline, inducing sparsity and efficient solution via MIP (Boutilier et al., 2019).

Mountainous/rural SAD system design incorporates full 3D travel-time cost functions $$\mathcal{L}_{bldg}(\theta) = \frac{1}{|\mathcal{B}_{hard}|} \sum_{i\in \mathcal{B}_{hard}} -[y_i\log \hat y_i + (1-y_i)\log(1-\hat y_i)].$$5 that consider vertical/horizontal drone speeds and terrain via Google Elevation API, enabling robust base allocation/assignment while comparing to historical helicopter response data (Wankmüller et al., 2019). Multi-objective variants balance fleet cost and mean/percentile response time, with backup/reshuffling for redundancy.

5. Field Performance and Empirical Results

Empirical evaluation includes both quantitative metrics and field-trial observations.

Subsystem	Precision (%)	Recall (%)	IoU (%)	Operating Latency	Reference
Road segmentation	76.48	70.96	58.08	0.80 s/MP (3.3 s/km²)	(Merkle et al., 2023)
Building segm.	83.74	77.70	68.12	0.38 s/MP (9.5 s/km²)	(Merkle et al., 2023)
Person detection	54.13	65.87	(AP:60.36)	0.44 s/MP (≈19 min/km²), 2 s/16 MP onboard	(Merkle et al., 2023)

Onboard, the person-detection CNN matches laptop performance with minimal loss in AP. Field trials validate the utility of remote sensing overlays for dynamic route planning, real-time hazard avoidance, and routing adaptation.

For AED delivery, optimization models show that modest numbers of drones (e.g., 1×3 base/drones for +1 min RT gain in urban, ≈2×2 in rural) can cut mean and 90th-percentile response times substantially, improving both efficiency (urban) and equity (rural) (Boutilier et al., 2019). In Alpine regions, optimally placed medium-range drones achieve mean response times of 2–5 min, 95% within 5 min—an order of magnitude improvement over helicopters (Wankmüller et al., 2019).

6. Safety, Security, and Fail-Safe Mechanisms

SAD systems implement layered safety:

• Real-time hazard detection for drop-zone clearance and mission aborts if humans are detected.
• Obstacle/collision avoidance via vision and active sensors (LiDAR, radar).
• Autonomous fallback in communication loss: onboard-only decision logic, RTH, and auto-land triggering on battery or GNSS faults (Mnaouer et al., 2021).
• Redundancy: dual sensors, watchdogs, and backup drone assignments (Wankmüller et al., 2019).
• Security: encrypted all-link telemetry (AES-256), mutual authentication, secure boot, and over-the-air update protocols (Mnaouer et al., 2021).
• Privacy: no identities stored, robust anonymization in logging (Merkle et al., 2023).

7. Situation Awareness Detection (SAD) for Human-Autonomy Teaming

Recent advances leverage real-time bystander situation awareness (SA) assessment using video, graph neural networks, and transformers. The DANDSD dataset captures high-frequency, multi-modal interactions for bystander-drone teaming in opioid-overdose simulations (Chang et al., 3 Oct 2025).

Operational SAD systems perform:

• Spatiotemporal graph construction (nodes: bystander, drone, instructor, victim), feature embedding with 2-layer GCNs, and sequence modeling via transformers (2 layers, 2 heads, d_model=32).
• Objective functions mix cross-entropy and IoU segmentation losses for binary and ternary SA prediction, with MoF and IoU for evaluation.
• Real-time inference enables adaptive drone guidance (e.g., highlighting victim location on low perception, audio cues on low comprehension), pipeline latencies <200 ms.
• Validated on DANDSD, transformer-based models achieve MoF=0.58, IoU=0.34, outperforming GCN/FINCH baselines by 9 and 5 percentage points, respectively (Chang et al., 3 Oct 2025).

A plausible implication is that integrating SAD with aid-delivery logic can further close the loop between assessment, guidance, and successful layperson intervention in out-of-hospital emergencies.

---

Drone-First-Aid SAD systems manifest as multi-layered, safety-critical, AI-augmented autonomy stacks for first response. Their architecture, control logic, optimization methodology, empirical validation, and safety practices provide a rigorous framework for deploying rapid, robust, and equitable life-saving interventions across diverse mission environments.