UAV Disaster Zone Rescue Priority Assessment: A Deep Feasibility Analysis for Final Year Projects

Project Concept: Assessing Rescue Priority from UAV Imagery

In a computer vision community on Reddit, a student team proposed a socially meaningful final-year project idea: using UAV aerial footage to automatically detect trapped individuals at disaster scenes and estimate rescue urgency based on a set of visually observable indicators.

The application of Computer Vision (CV) in disaster response can be traced back to search-and-rescue robotics research in the early 2000s. With the rise of deep learning — particularly AlexNet's breakthrough performance at the 2012 ImageNet competition — neural network-based object detection accuracy improved dramatically, moving UAV-based automated search-and-rescue systems from theory into engineering practice. However, the core challenges of disaster-scene CV are not purely algorithmic; they are fundamentally about data. The massive annotated datasets accumulated by industry largely come from routine scenarios like traffic surveillance and autonomous driving, creating a fundamental domain shift problem when applied to disaster environments — and this is the primary engineering challenge the team faces.

To their credit, the team maintained a clear-eyed understanding of project boundaries from the start. They explicitly noted that this system is not a medical triage system — genuine injury assessment requires thermal imaging, vital-sign sensors, and other hardware they simply don't have access to.

Medical Triage derives from the French word "trier" (to sort) and is a standardized emergency medicine process for allocating limited medical resources based on injury severity. Widely used field triage systems include START (Simple Triage and Rapid Treatment) and SALT (Sort-Assess-Lifesave-Treat/Transport), with core assessment metrics covering respiratory rate, capillary refill time, and level of consciousness. Obtaining these parameters requires thermal imaging (body temperature distribution), millimeter-wave radar (respiration detection), or direct physiological sensor contact — all of which are well beyond the capabilities of standard optical cameras. The team's explicit distinction between "visually observable priority" and "medical triage" reflects a clear understanding of hardware capability limits and avoids the ethical risk of the system being misused for medical decision-making.

Instead, they focused on purely visually observable signals:

• Posture recognition: standing, prone, crawling, and other body positions
• Prolonged immobility: using temporal analysis to determine whether an individual remains stationary over time
• Proximity to hazards: distance from fires, floodwaters, rubble, and other dangers

These signals would ultimately be weighted and fused into a "priority score" for each detected individual, generating a ranked list for rescue teams. This is a structurally clean engineering proposition — and an ethically sensitive one.

The Core Bottleneck: A Severe Shortage of Disaster-Scene Aerial Datasets

The team itself identified the biggest obstacle: a lack of publicly available datasets matching this specific scenario. This is precisely the reason many CV final-year projects fail.

Why General Datasets Can't Be Directly Reused

Aerial human detection in disaster scenes has several unique characteristics that make off-the-shelf datasets difficult to transfer:

1. Significant perspective difference: The top-down and oblique angles of UAV imagery differ drastically from ground-level cameras and standard pedestrian detection datasets, severely limiting model generalization.
2. Scarcity of non-standard posture samples: Most everyday datasets show people standing or walking, while poses like prone, partially buried, or crawling are nearly absent from public datasets.
3. Complex environmental interference: Smoke occlusion, water surface reflections, and debris all significantly degrade detection accuracy.

While datasets like VisDrone, SARD (Search and Rescue Dataset), and HERIDAL exist specifically for aerial search-and-rescue scenarios, most only annotate "human location" — they lack fine-grained labels for posture category, duration of stillness, or distance to hazards, leaving a significant gap from the team's actual needs.

A Closer Look at the Three Key Datasets: VisDrone, developed primarily by Tianjin University, is a large-scale UAV visual dataset containing over 10,000 images and 4 million annotated bounding boxes covering 10 object categories including pedestrians and vehicles — but its scenes are predominantly urban daily environments, lacking the environmental interference characteristic of disasters. HERIDAL (Helicopter Rescue Image Dataset for Person Detection in Alpine Landscapes) is designed specifically for mountain search-and-rescue, with images from helicopter aerial footage and human detection annotations in complex terrain, making it one of the closest publicly available datasets to real rescue scenarios. SARD focuses on human detection from a UAV perspective and provides some posture variation samples, but its data volume is relatively limited. All three share a common limitation: annotation granularity stops at the bounding box level, with none including semantic-level labels for posture categories, temporal stillness states, or hazard proximity relationships.

Feasibility Assessment: Layered Execution — Scope Convergence Is Key

From an engineering implementation standpoint, this project is "doable" — but doing everything within the time and resource constraints of a final-year project is not realistic. Here is a layered breakdown of a practical implementation path.

Layer 1: Aerial Human Detection (Highest Priority, Most Feasible)

This is the foundation of the entire system pipeline and the most technically mature component. Fine-tuning a YOLO-series or DETR-class detector on the SARD/HERIDAL datasets can deliver a reliable detection model within a typical final-year project timeline. This should be treated as the core deliverable, establishing a solid foundation before considering additional modules.

Technical Differences Between YOLO and DETR: The YOLO (You Only Look Once) series has undergone multiple iterations from YOLOv1 to YOLOv10 since its 2016 debut. Its core idea is to unify object detection as a single-pass regression problem — dividing the image into grid cells and predicting bounding boxes and class probabilities within each cell, achieving very high inference speeds (100+ FPS) well-suited for real-time processing on UAV edge computing devices. DETR (Detection Transformer), proposed by Facebook AI Research in 2020, was the first to bring the Transformer architecture into object detection, directly predicting detection results in a sequence-to-sequence manner without hand-designed anchors or non-maximum suppression (NMS) post-processing, delivering more stable performance in complex occlusion scenarios at the cost of higher computation and slower training convergence. For resource-constrained final-year project settings, YOLOv8 or YOLOv9 is generally the preferred choice for balancing accuracy and deployment efficiency.

Layer 2: Posture Recognition and Stillness Detection (Moderate Difficulty — Recommended as an Extension Module)

Posture classification can be added as a lightweight classification head on top of detection bounding boxes, but the challenge is that effective training data is nearly nonexistent. Two practical paths exist:

• Synthetic data augmentation: Use simulation platforms like Unreal Engine or AirSim to generate synthetic aerial images with posture labels, expanding samples for non-standard poses like prone and crawling.

Simulation Platforms and Synthetic Data: AirSim is Microsoft's open-source UAV simulation platform built on Unreal Engine, natively supporting multiple sensor simulations (RGB, depth, infrared) and physically accurate flight dynamics. It can generate aerial images with precise pixel-level annotations and has been used in multiple search-and-rescue CV studies to supplement training data. The core challenge with synthetic data is the "Reality Gap" — non-negligible differences between simulated rendering and real-world textures and noise characteristics. This is typically addressed using Domain Randomization (randomly varying lighting, textures, and weather in simulation to force the model to learn rendering-agnostic robust features) or Domain Adaptation (using adversarial training or feature alignment to bring simulated and real-world feature distributions closer together).

• Stillness duration detection: This relies on multi-frame video sequences rather than single images and can be implemented using a lightweight object tracking algorithm (such as ByteTrack) combined with a displacement threshold — considerably more manageable in engineering complexity than posture classification.

ByteTrack Technical Overview: Multi-Object Tracking (MOT) is the technology of continuously associating detection results with the same physical individual across video sequences. ByteTrack, proposed by Megvii in 2021, innovates by departing from conventional tracking methods that only associate high-confidence detection boxes — ByteTrack's two-stage matching strategy incorporates low-confidence detection boxes as association candidates as well, significantly reducing ID switches in occlusion and brief disappearance scenarios. For stillness detection, ByteTrack maintains a unique ID and position trajectory across frames for each detected individual, determining whether an individual is continuously stationary by checking whether centroid displacement across N consecutive frames falls below a preset threshold. This approach is significantly simpler to implement than end-to-end posture classification and is the recommended implementation path for the stillness detection module.

Layer 3: Priority Weighted Scoring (Highest Ethical Risk — Needs to Be Repositioned)

This is the part of the project with the most questionable scientific validity and highest ethical risk. Combining posture, stillness, and hazard proximity into a single "rescue priority" score essentially means building a decision model with no ground truth — the system has no way to validate against real annotations whether the scores are correct.

Recommended repositioning: Rather than framing this as "the system automatically determines rescue order," explicitly position it as an assistive information visualization tool that presents individual reference indicators to rescue personnel, with final decision-making authority remaining with humans.

AI-Assisted Decision-Making Ethics Background: Positioning AI systems as "assistive tools" rather than "autonomous decision-makers" is one of the core consensuses in current AI ethics. The EU AI Act classifies high-risk AI systems affecting citizen safety — including emergency response scenarios — as a strictly regulated category, requiring explainability, human oversight mechanisms, and complete error-handling procedures. In disaster response, both FEMA (Federal Emergency Management Agency) and INSARAG (International Search and Rescue Advisory Group) emphasize that technical assistance tools must preserve a "Human-in-the-Loop" decision structure. From a technical implementation standpoint, explainability requirements encourage the team to choose transparent rule-based weighting rather than a black-box neural network for generating priority scores — every output value in the former can be traced back to specific visual observation indicators, satisfying the basic requirement of auditability and making it easier to provide technical evidence in post-disaster accountability reviews.

This positioning both fits engineering reality and effectively mitigates ethical controversy.

Practical Recommendations for the Project Team

The key to this final-year project is not "whether all planned features can be implemented," but rather how to reasonably converge scope and honestly communicate system capability boundaries.

Scope Convergence Strategy

• Solidify the main track: Aerial human detection + semantic segmentation of hazard sources (fire/flood/rubble all have relatively mature segmentation methods). These two components have reliable data support and should serve as the project backbone.
• Treat posture and stillness modules as experimental extensions: Explicitly label them as proof-of-concept stage. Validating the approach with synthetic data is sufficient — production-level accuracy is not required.
• Convert the scoring system into explainable visualization: Use transparent rule-based weighting rather than a black-box model, ensuring every score can be traced back to specific visual observation evidence.

Data Acquisition Strategy

• Prioritize public datasets like SARD, HERIDAL, and VisDrone as the training foundation.
• Supplement scarce non-standard posture samples using simulation platforms like AirSim or CARLA.
• Use small amounts of real-scene data for validation set evaluation only — not for model training.

Ethical Statements and Paper Framing

In both the thesis and defense, proactively stating system limitations is critical: this system does not perform medical triage, does not replace human judgment, and all outputs are for reference only. This kind of transparency won't lower your score — it will actually serve as a demonstration of the team's engineering maturity.

Conclusion

The value of this project does not lie in whether it could become a production system actually deployed in disaster zones — at the resource scale of a final-year project, that is essentially out of scope. Its real value lies in demonstrating how a team makes well-documented engineering decisions under real-world constraints of scarce data, sensitive requirements, and technical limitations.

Breaking down an ambitious vision into "deliverable core components" and "experimental extensions," honestly confronting dataset gaps and ethical boundaries — this itself is one of the most fundamental skills in computer vision engineering practice. A solid aerial detection model paired with an honest declaration of capability limits is far more persuasive — and far more commendable — than chasing an overambitious "all-capable rescue AI."