Coal Face AI Tests Edge Drone Vision to Cut Man-Overboard Deaths

Panel 3 - 2: AI-Driven Person Detection for Maritime Search and Rescue: The Value of Product-Based — Ai4

Ai4

Organizations waste millions by buying AI models before establishing operational need, whereas empirically validating the single riskiest assumption on low-cost hardware prevents failed deployments.

AI-assisted summary of Panel 3 - 2: AI-Driven Person Detection for Maritime Search and Rescue: The Value of Product-Based — original source. Check the original for full context.

Source published: 2026-08-14

Detailed notes

## Background Published on YouTube by Ai4 on August 14, 2026, this panel presentation features Thomas Wilson, founder of Coal Face AI and a 21-year Royal Navy veteran. The talk uses his master's research into AI-driven maritime search and rescue to demonstrate how grounding AI development in deeply understood operational problems prevents organizations from chasing shiny objects and building the wrong products.

## TL;DR Wilton argues against the pervasive 'technology-first' trap—in which organizations acquire AI models and then search for problems to solve—advocating instead for a disciplined product-based research framework centered on validating the single riskiest assumption first. Grounded in the stark reality that over 40% of maritime man-overboard incidents end in fatalities, he details the development of Vantage, an AI system that pairs drone thermal imaging with computer vision for rapid ocean casualty detection. By collecting and manually labeling over 10,000 thermal frames in freezing field conditions at Manvers Lake, the project achieved 97% precision and 96% recall while executing locally on commodity edge hardware. He concludes that sustainable AI adoption requires clear UI/UX context, edge feasibility, and designing strictly around human operators in the loop rather than autonomous hype.

## Atmosphere & style The presentation carries a grounded, candid, and refreshingly pragmatic tone that cuts through Silicon Valley buzzwords with no-nonsense military realism. Wilton speaks with the quiet authority of a seasoned naval operator, blending dry humor about defense bureaucracy asking for 'magic AI dust' with visceral recollections of staring off pitch-black flight decks into freezing waters. Watching the talk feels like receiving an urgent, hard-won briefing from an operational pragmatist who cares far more about saving a drowning casualty than flattering executive tech enthusiasm.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=4s) Organizations frequently fall into the technology-first trap, buying AI models or capabilities before proving an acute operational need exists. - [03:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=185s) Asking 'How can we use AI?' leads to wasteful shiny-object adoption; teams should instead ask if a specific AI capability solves an authentic, validated problem. - [05:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=320s) The product-based research framework focuses on identifying frontline operational friction, formulating testable product hypotheses, and isolating the single riskiest assumption. - [05:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=320s) The greatest economic value of applied AI research is preventing organizations from spending millions building the wrong product, rather than merely validating initial intuition. - [06:53](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=413s) Over 40% of maritime man-overboard incidents result in loss of life, with the primary bottleneck being the human visual difficulty of spotting a person in open water. - [09:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=545s) Empirical testing under authentic environmental stressors—such as freezing winter water and real human swimmers—is essential to validate core AI detection assumptions. - [12:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=740s) Manual data annotation and analyzing diminishing returns in dataset scaling help teams optimize resource allocation when collecting expensive domain data. - [12:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=740s) Maritime search and rescue AI must execute inference locally on commodity edge hardware to eliminate dependence on cloud connectivity in remote ocean environments. - [12:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=740s) Computer vision models achieved 97% precision and 96% recall in detecting open-water casualties on low-cost edge platforms. - [14:25](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=865s) AI systems must be designed as assistive tools that provide clarity, speed, trust, and context, keeping human operators in control of life-or-death decisions. - [16:08](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=968s) Transitioning research prototypes to production requires testing varied sea states and integrating thermal computer vision with radar, GPS, and AIS navigation feeds. - [17:17](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=1037s) Adoption barriers regarding confidence, cost, and trust are overcome by starting with the operator's problem, deploying to low-cost hardware, and providing transparent model context.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=4s) Introduction: Coal Face AI and Shifting from Technology to Problem

Tom Wilton opens the session by introducing his background and his venture, Coal Face AI. Deeply rooted in the industrial coal-mining heritage of South Yorkshire in the UK, the company derives its name from its mission: delivering artificial intelligence where it matters most, directly at the coal face. Before founding the company, Wilton served for just over 21 years in the Royal Navy, concluding his military career in the Navy's AI transformation cell. Having spent countless years deployed at sea and working directly alongside maritime search and rescue organizations, he witnessed firsthand the immense operational difficulty of spotting and recovering an individual lost in the vast, open ocean.

His master's research, titled around AI-driven person detection for maritime search and rescue utilizing YOLO object detection, initially looked like a standard engineering project combining drones, thermal imaging, and computer vision. However, during the course of the project, Wilton realized that the algorithms and hardware were not the most critical or interesting takeaways. The truly transformative revelation was understanding why the solution was built in the first place. That shift in perspective reshaped how organizations ought to approach AI adoption. Rather than simply showcasing technical benchmarks, the goal of his talk is to outline a mindset that can prevent organizations from burning through millions of dollars and years of wasted effort chasing hype.

- **Background:** Founder of Coal Face AI, veteran with 21+ years in the Royal Navy, formerly in the Royal Navy AI transformation cell. - **The Core AI Dilemma:** Every week delivers a new model release, monthly AI strategies, and endless conference demos, yet organizations continuously fail to extract tangible business value. - **The Anti-Pattern:** "They've bought the hammer, and now they're desperately looking for nails."

---

[03:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=185s) The Technology-First Trap and "Magic AI Dust"

Whenever enterprise leadership and public organizations discuss artificial intelligence, the conversation almost universally starts with the same question: "How can we use AI?" While this inquiry sounds rational on the surface, Wilton argues it is one of the most financially ruinous questions an organization can ever ask. It assumes by default that AI inherently belongs somewhere in the tech stack, forcing adoption from the top down. The far more rigorous, value-driven question is: "Does this specific AI capability solve a real-world problem, and can we demonstrate that before we invest?" That subtle reframing forces AI to earn its place inside an organization based on merit rather than executive enthusiasm.

To illustrate how pervasive this upside-down thinking has become, Wilton recounts a memorable exchange from his time in the Navy's AI transformation cell. Representatives from the UK's Submarine Delivery Agency approached him while designing the operating system for the Dreadnought class—the nation's next-generation nuclear deterrent ballistic missile submarines. Their inquiry was simply: "What AI can I put in it? What magic AI dust can I sprinkle in there?" Instead of identifying an acute operational hurdle and assessing whether AI offered an effective remedy, the desire was simply to check an AI box on a flagship defense program. Wilton likens this behavior to a magpie drawn to glittering trinkets: "AI is today's shiniest object. But organizations don't fail because they bought AI. They fail because they bought AI before proving that they actually needed it."

```mermaid flowchart TD accTitle: Technology-First vs Problem-First AI Adoption subgraph Bad["Flawed Approach: Shiny Object Adoption"] B1["Buy AI / Model First"] --> B2["Hunt for Internal Problems"] B2 --> B3["Costly Integrations & Unused Tools"] end subgraph Good["Product-Based Research Framework"] G1["Identify Operational Problem"] --> G2["Form Product Hypothesis"] G2 --> G3["Isolate Riskiest Assumption"] G3 --> G4["Low-Cost Empirical Test"] G4 --> G5["Evidence-Driven Investment"] end ```

---

[05:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=320s) The Product-Based Research Framework

To combat shiny-object syndrome, Wilton presents an alternative disciplined methodology called "product-based research." The framework begins not with an open software license, a vendor demo, or an executive returning inspired from a tech conference, but with an authentic, painful operational bottleneck experienced by frontline human operators. From this validated friction, teams formulate a product hypothesis: "I believe this capability might solve this specific problem." Crucially, a real hypothesis must welcome the possibility of being proven completely wrong.

The next step demands isolating the single riskiest assumption underlying the entire concept. In Wilton's research, the critical unknown was neither whether drones could achieve stable flight, nor whether commercial thermal cameras existed, nor whether the YOLO architecture was sound. The single existential question was: *Can AI detect a human in open water reliably enough to be operationally useful?* If that assumption fails, the product is dead on arrival, meaning teams should never spend capital building full end-to-end architectures before validating that core unknown. Wilton emphasizes that rigorous research does not exist to flatter egos and prove initial intuition right; its greatest economic value lies in preventing organizations from wasting years building the wrong product.

- **Real-World Problem:** Anchor strictly to an observable operational pain point. - **Product Hypothesis:** Define a testable assertion linking a specific capability to a specific outcome. - **Riskiest Assumption:** Isolate the single failure point that could completely invalidate the concept. - **Empirical Validation:** Test the assumption under raw, representative conditions before writing production software. - **Evidence-Driven Investment:** Require each experimental milestone to earn the subsequent tranche of funding.

---

[06:53](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=413s) The Reality of Maritime Search and Rescue

Drawing directly from his 21 years at sea, Wilton grounds the framework in the visceral realities of naval operations. Royal Navy regulations permit off-watch personnel to consume alcohol during downtime, and Wilton recalls standing on the pitch-black upper decks of warships at night—where exterior deck lighting is strictly absent—staring into the dark water and wondering: "If I fell in, is anyone coming? Did anyone see me fall in? And if they did, are they going to reach me in time?"

This hazard is backed by grim statistics. Between 2015 and 2023, the UK registered over 300 reported man overboard incidents. In the United States, 2023 alone saw more than 200 occurrences. Most devastatingly, over 40% of all recorded man overboard incidents result in loss of life. When surveying the maritime domain to understand why modern AI adoption remains stubbornly low, three recurring barriers emerged: lack of confidence in AI systems, the daunting perceived cost of integrating AI into complex shipboard hardware, and a lack of trust from deck operators. Furthermore, experienced search and rescue personnel confirmed that the primary operational bottleneck is not navigation, fleet communication, or multi-asset coordination; it is the sheer physical difficulty of spotting a human head bobbing in open water.

---

[09:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=545s) Designing the Experiment at Manvers Lake

Armed with the problem, Wilton defined the product hypothesis: a drone-mounted thermal camera paired with a computer vision object detection model can reliably locate a person in distress at sea faster and under conditions where human visual scanning degrades. Because the riskiest assumption was whether the model would miss targets in real water, testing required authentic environmental stressors: true thermal sensors, wild water, cold weather, and living human subjects.

The field trials were conducted roughly three miles from Wilton's home at Manvers Lake in South Yorkshire, a body of water reclaimed from an old colliery site. In February, under freezing conditions with a surface temperature of -2°C (translated with a brief aside for American listeners as 28.4°F), local open-water swimmers volunteered to enter the freezing lake for drone thermal passes. Over 70 minutes of raw thermal video were captured from the aerial platform.

Rejecting automated shortcut tools in favor of creating a gold-standard, trustworthy dataset, Wilton hand-labeled every single frame himself, generating over 10,000 annotated images. The resulting dataset captured varied target signatures: lone human swimmers, humans equipped with flotation devices, background distractors like abundant ducks and swans, and even a waterside igloo sauna on the jetty. When evaluating initial model training performance across dataset sizes, training on an initial batch of 50 images yielded poor detection, whereas scaling to 1,000 images sparked a dramatic surge in accuracy, laying the empirical groundwork for how data volume directly drives model utility.

- **Trial Setting:** Manvers Lake (reclaimed colliery site, South Yorkshire, UK) in February winter conditions (-2°C / 28.4°F). - **Data Capture:** Over 70 minutes of continuous aerial thermal drone footage. - **Manual Annotation:** 10,000+ hand-labeled frames to guarantee data integrity and eliminate label hallucination. - **Dataset Diversity:** Open-water swimmers, flotation aids, false-positive distractors (wildlife such as swans and ducks), and thermal anomalies (igloo sauna on the jetty). - **Initial Training Progression:** Poor accuracy at 50 images, followed by steep, dramatic performance gains at the 1,000-image threshold.

[12:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=740s) Evaluation Metrics, Diminishing Returns, and Edge Inference

When analyzing the scaling of the training dataset, doubling the dataset size yielded an initial performance leap, but subsequent doublings produced significantly smaller gains. From an engineering standpoint, identifying this point of diminishing returns is vital because collecting and annotating maritime data is labor-intensive and expensive. Recognizing where additional financial and time investment begins to offer progressively smaller benefits allows teams to budget resources effectively.

```mermaid flowchart LR accTitle: Maritime SAR Edge AI Pipeline subgraph DataEdge["Edge AI Pipeline"] A["Thermal Sensor Feed"] --> B["Jetson Xavier / Edge Unit"] B --> C["Object Detection Model"] end subgraph OutputMetrics["Operator Delivery"] C --> D["Confidence & Context"] D --> E["Human Operator Decision"] end ```

The model's performance was evaluated across standard computer vision metrics. Rather than focusing purely on abstract statistics, each metric corresponds to an operational reality: precision answers how often the system is correct when it claims to have detected someone, recall measures how many people actually in the water were successfully spotted, and the F1 score provides a balanced composite of the two. Mean average precision (mAP) scores confirmed the accuracy of the bounding box placements. Finding people, however, is only half the battle; latency and deployability dictate operational survival. Because vessels operating in remote environments—such as the middle of the Atlantic or Arctic Ocean—lack reliable high-bandwidth connectivity, cloud processing is unviable. The inference must occur locally on commodity edge hardware, specifically tested on platforms like the NVIDIA Jetson Xavier. The experimental results confirmed the core hypothesis:

- **Precision:** 97% - **Recall:** 96% - **F1 Score:** 96% - **Edge Feasibility:** Low-cost edge devices deliver real-time inference without cloud connectivity, establishing that the core detection mechanism is credible enough to justify further development.

[14:25](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=865s) Human-Centric UI/UX and the Concept of Operations

Validating the algorithm does not guarantee operational adoption. Building an effective system requires addressing how human operators interact with the output under extreme stress. Four core pillars define this interface: clarity (understanding the AI output instantly), speed (enabling decisions in seconds rather than minutes), trust (believing recommendations enough to take immediate action), and context (understanding *why* the AI flagged a specific target). The system is deliberately designed not as an autonomous replacement, but as an assistive tool where human judgment remains firmly at the center of the loop.

The operational reality in modern maritime services highlights the need for technological intervention. When a man-overboard alarm sounds—even on modern Royal Navy vessels—standard procedure still relies on shipwide personnel rushing to the upper deck to scour the ocean with binoculars and the naked eye. An AI-assisted Concept of Operations (CONOPS) introduces a structured, automated response:

```mermaid flowchart TD accTitle: Man Overboard Automated Response Workflow subgraph AlertPhase["Alert & Dispatch"] A["Man Overboard Alert Triggered"] --> B["Automated Drone Launch"] B --> C["Autonomous Search Path Execution"] end subgraph DetectionPhase["Detection & Navigation"] C --> D["Thermal Detection & Tracking"] D --> E["Operator Alert + Course-to-Steer Recommendation"] end ```

- When an incident occurs, a drone launches immediately and calculates an optimal search trajectory. - The drone scans the water, detects the casualty using real-time computer vision, alerts the bridge, and calculates an exact course to steer for recovery.

[16:08](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=968s) Research Limitations and Roadmap for Maturation

Rigorous research requires transparency regarding boundary conditions and what has not yet been proven. The initial project did not evaluate detection performance across violent sea states, severe meteorological obstructions (such as fog, heavy haze, and sea frets), or diverse sensor payloads and airframes. It is not intended as an immediate, turnkey replacement for existing naval search-and-rescue drills.

To advance the system from a research prototype to an operational capability, three development milestones must be met:

- **Dedicated Edge Validation:** Transitioning the model from simulated compute benchmarks onto hardened, flight-ready edge hardware payloads. - **Dataset Diversification:** Expanding the data repository not merely in raw volume, but across environmental variables—including differing sea states, ambient water temperatures, salinity levels, diurnal cycles, lighting conditions, and diverse human subjects. - **Sensor Fusion:** Integrating thermal video feeds with onboard navigational and radar systems, including Radar, GPS, and Automatic Identification Systems (AIS), so multiple data streams synthesize into a unified operational picture for the operator.

[17:17](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=1037s) Overcoming Adoption Barriers and the Product-First Framework

Surveys of maritime search-and-rescue stakeholders revealed three recurring barriers preventing the operational adoption of AI: confidence, cost, and trust. The project was explicitly structured to dismantle each hurdle:

The resulting system, named **Vantage**, demonstrates that an organization can design operationally viable, edge-deployable, life-saving AI starting from zero domain data by adhering strictly to product-first principles.

Rather than following the common tech-industry anti-pattern of building an arbitrary capability and searching for a problem to apply it to, practitioners should invert the paradigm: "Start with the person in the water and everything else follows." Organizations must first deeply understand the operational problem, systematically validate the riskiest assumptions, engineer for the physical environment of deployment, and only then commit capital to scaled production.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=4s) They've bought the hammer, and now they're desperately looking for nails. — Tom Wilton

> [03:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=185s) How can we use AI? — Tom Wilton

> [03:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=185s) Does this specific AI capability solve a real-world problem, and can we demonstrate that before we invest? — Tom Wilton

> [03:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=185s) What AI can I put in it? What magic AI dust can I sprinkle in there? — Tom Wilton

> [03:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=185s) AI is today's shiniest object. But organizations don't fail because they bought AI. They fail because they bought AI before proving that they actually needed it. — Tom Wilton

> [05:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=320s) I believe this capability might solve this specific problem. — Tom Wilton

> [06:53](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=413s) If I fell in, is anyone coming? Did anyone see me fall in? And if they did, are they going to reach me in time? — Tom Wilton

> [17:17](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=1037s) Start with the person in the water and everything else follows. — Tom Wilton

## Mentioned Tom Wilton, Coal Face AI, Royal Navy, Royal Navy AI transformation cell, Submarine Delivery Agency, Dreadnought class, YOLO, Manvers Lake, NVIDIA Jetson Xavier, Automatic Identification Systems (AIS), GPS, Radar, Vantage

## Video description Thomas Wilson, Founder, Coal Face AI

What if the greatest value of AI research is not proving that a product will work, but preventing you from building the wrong one? Drawing on more than 20 years of Royal Navy operational experience, I will use my research into AI-driven person detection for maritime search and rescue to show how a deeply understood real-world problem can be translated into a product hypothesis and tested before major investment is committed. The talk will explore how organisations can identify the riskiest assumptions, choose meaningful measures of success and distinguish a promising model from a genuinely valuable product. Attendees will leave with a practical, problem-led framework for moving from research to prototype, trial and informed AI investment. Join this session to learn how to avoid shiny-object adoption and make evidence, rather than enthusiasm, the basis for your next AI decision.

[Source](https://www.youtube.com/watch?v=N8_cPkXcAiI)

中文简报

Coal Face AI测试无人机边缘视觉技术以减少落水死亡事故许多机构在确立实际业务需求前便采购AI模型造成数百万资金浪费，而在低成本硬件上实证验证单一最高风险假设则可避免部署失败。 Wilton对普遍存在的“技术先行”陷阱提出了质疑——即机构往往先采购AI模型再去寻找要解决的问题——转而主张采用严密且以产品为导向的研究框架，其核心在于优先验证单一最高风险的假设。基于海上人员落水事故死亡率超过40%的严峻现实，他详细阐述了Vantage的研发过程：该AI系统将无人机热成像与计算机视觉相结合，用于快速搜寻海上落水人员。通过在Manvers Lake严寒的实地环境下采集并手动标注超过10,000帧热成像画面，该项目在通用边缘硬件上实现了本地运行，并达到了97%的精确率与96%的召回率。他总结道，可持续的AI落地应用需要清晰的UI/UX环境、边缘端可行性，并且必须严格围绕“人在回路”的操作人员进行设计，而非盲目追逐自主化的噱头。许多组织常陷入“技术优先”的追逐热点误区，而基于产品的研究框架主张从一线业务痛点出发，在投入巨资前先隔离并低成本验证最关键的高风险假设，以防构建错误的产品。 (3:05) 超过40%的海上人员落水事故会导致丧生，搜救的核心痛点在于仅凭人眼在广阔、黑暗且冰冷的海域中搜寻人体目标的生理与视觉极限。 (6:53) 在-2°C低温的曼弗斯湖实地测试中，基于无人机热成像采集与逾万张手工标注样本，YOLO目标检测模型达到了97%的精确率和96%的召回率，同时验证了数据扩展越过千张阈值后边际效应递减的规律。 (9:05) 由于远海无法依赖云端网络连接，海难搜救AI必须依托低成本通用边缘硬件（如NVIDIA Jetson Xavier）实现本地实时推理，并以人机协作界面破除行业对AI信任度与集成成本的顾虑。 (12:20)

Reading summary

Tom Wilton, founder of Coal Face AI and a Royal Navy veteran, argues that AI initiatives frequently fail because organizations purchase technology before identifying real operational needs. Drawing on his prototype for maritime search and rescue, he outlines a product-first research framework that validates core assumptions under authentic field conditions prior to software development. The Technology-First Trap and Product-Based Methodology (3:05) Wilton states that organizations routinely waste millions of dollars asking 'How can we use AI?'—a question that treats technology as an automatic requirement rather than a tool earning its place. During his tenure in the Royal Navy's AI transformation cell, defense planners for the Dreadnought-class submarine asked him what 'magic AI dust' could be sprinkled into their operating systems, demonstrating a widespread tendency to seek problems for pre-selected technologies. To counter this anti-pattern, Wilton proposes a product-based research framework. Teams must begin with a documented operational bottleneck experienced by frontline operators and isolate the single riskiest assumption that could invalidate the entire project. Each development phase and tranche of capital must be earned through low-cost, empirical tests under raw conditions before engineers build end-to-end architectures. Maritime Search Bottlenecks and Empirical Testing (6:53) Grounding his thesis in maritime search and rescue (SAR), Wilton cites over 300 reported man-overboard incidents in the UK between 2015 and 2023, more than 200 in the US in 2023 alone, and a fatality rate exceeding 40%. Operator interviews confirmed that the primary operational bottleneck is not fleet communications or navigation, but the physical difficulty of visually spotting a human in open water. His research hypothesis centered on whether computer vision combined with drone-mounted thermal imaging could reliably detect casualties faster than visual scanning. To test this core assumption, field trials were staged at Manvers Lake in freezing conditions (-2°C / 28.4°F) using volunteer open-water swimmers. Rather than using automated annotation shortcuts, Wilton manually labeled over 10,000 frames from 70 minutes of aerial thermal video. The dataset incorporated swimmers, flotation aids, and distractors such as waterfowl and a lakeside sauna. While initial training on 50 images yielded poor accuracy, scaling to 1,000 images triggered a sharp rise in model utility before hitting diminishing returns. Edge Validation, Operator Trust, and Human-in-the-Loop Integration (12:20) Because maritime vessels often operate in remote environments without high-bandwidth cloud access, Wilton targeted real-time inference on commodity edge devices such as the NVIDIA Jetson Xavier. The resulting YOLO-based model achieved 97% precision, 96% recall, and a 96% F1 score, demonstrating edge feasibility for real-time operations. Wilton stresses that algorithmic accuracy alone does not guarantee operational adoption, which has historically been stymied by lack of operator confidence, high perceived integration costs, and low trust in black-box systems. The resulting prototype, named Vantage, was integrated into a proposed Concept of Operations where an automated drone detects a casualty and suggests a recovery course. However, human judgment remains strictly central: the user interface exposes confidence scores and contextual metadata so operators understand why a detection was flagged. Wilton acknowledges several research limitations: the model has not yet been validated across extreme sea states, dense fog, varying water salinity, or diverse sensor payloads. Fully maturing the capability will require testing flight-hardened edge payloads and fusing thermal imagery with onboard Radar, GPS, and AIS navigation feeds.

精炼阅读摘要

Coal Face AI 创始人 Tom Wilton 提出，组织在引入人工智能时应当摒弃盲目追逐炒作的“技术先行”陷阱，转向以一线作战痛点为核心的“产品化研究框架”。通过海上搜救落水人员检测系统 Vantage 的研发历程，他证明了在低成本边缘设备上直接验证最高风险假设，能够在避免资本浪费的同时提供可靠的生命救援支持。警惕“AI魔法粉尘”与产品化研究框架 (3:05) Wilton 指出，政企机构在涉足人工智能时常陷入一种反模式：过早购入算法这把“锤子”，随后四处寻找能敲击的“钉子”。他在英国皇家海军 AI 转型部门任职期间，潜艇交付局在为下一代“无畏”级（Dreadnought class）弹道导弹核潜艇设计操作系统时，曾直接询问可以向系统中“撒入何种 AI 魔法粉尘”，而非评估具体作战瓶颈。Wilton 强调，“如何使用 AI”是一个极具财务破坏性的问题；组织并非因采购 AI 而失败，而是因在证明实际业务需求前就盲目买入而招致沉没成本。为解决这一困境，Wilton 提出了“产品化研究框架”（Product-Based Research Framework）。该方法要求摒弃以技术厂商演示为起点的路径，严格从一线人员的实际业务摩擦出发提出可证伪的产品假设。团队应进一步剥离次要因素，精准锁定足以颠覆整个设想的“最高风险假设”，并在高代表性环境下以最低成本开展实证验证。唯有假设得到数据支持，组织才应追加阶段性投资，从而使 AI 凭实际成效在机构内部赢得立足之地。海上搜救严酷现实与孤立核心风险假设 (6:53) 结合在英国皇家海军服役逾21年的深海经历，Wilton 阐述了海上搜救（SAR）面临的严酷现实。官方数据显示，2015至2023年间英国记录了超过300起人员落水事件，美国仅在2023年就发生了200多起，且超过40%的落水事件导致人员死亡。资深搜救人员证实，制约救援效率的核心瓶颈并非船只航行或通信协同，而是在大浪翻滚、无甲板照明或严寒环境下，人眼肉眼搜寻起伏漂浮的人头极易陷入视觉疲劳和视距盲区。尽管痛点明确，海上领域对 AI 的采纳却因缺乏系统信心、船载集成成本高昂以及基层信任不足而严重受阻。针对将无人机热成像与 YOLO 目标检测模型结合的构想，Wilton 认定工程团队无需优先关注无人机飞行姿态或商用红外相机的现有成熟度，而应集中资源攻关唯一的生存性假设：AI 系统能否在真实开放水域中足够可靠地识别人体以具备战术价值。若该假设破产，整体系统开发便应立即止损。曼弗斯湖实地测试与低成本边缘端验证 (9:05) 为在严苛真实条件下测试关键假设，研发团队在英国南约克郡曼弗斯湖（Manvers Lake，煤矿改造水域）展开实地冬季测试。在水表温度达 -2°C（28.4°F）的环境下，志愿游泳者下水配合无人机采集了逾70分钟的热成像视频。为规避自动化工具带来的“标注幻觉”并确立黄金基准，Wilton 独立手工标注了超10,000张图像，涵盖单独泳者、救生装备、天鹅野鸭等野生动物干扰源，以及岸边桑拿房等热异常干扰。实验表明，训练数据在从50张增至1,000张时模型精度显著跃升，但随后的倍增呈现出边际收益递减，这为高成本海洋数据采集提供了量化预算参考。在性能指标上，系统最终达到 97% 的精确率、96% 的召回率与 96% 的 F1 分数。为解决远洋公海无高带宽网络的问题，该模型成功部署在 NVIDIA Jetson Xavier 等低成本商用边缘计算硬件上实现本地实时推理，无需依赖云端，充分验证了技术可行性。人机协同作战概念与拆解采纳壁垒 (14:25) 算法精度并不自动转化为实战战力。Wilton 强调，必须围绕清晰度、决策速度、信任度以及判定上下文四大支柱设计交互界面，确保系统作为辅助工具赋能操作员，而非让黑盒算法剥夺人类判断。针对现代舰船在拉响落水警报后仍普遍依靠全舰人员携望远镜在甲板肉眼瞭望的现状，研究构建了新型作战概念（CONOPS）：警报触发后无人机自主升空并执行最优搜索路径，一旦热成像模型锁定目标，系统在向舰桥报警的同时直接解算回传救援转向航向（Course-to-Steer），将关键决策耗时缩短至数秒。这套名为 Vantage 的系统证明，坚持产品优先方法能有效击破行业采纳壁垒：以真实飞行实地数据替代合成数据以确立信心，以边缘硬件摆脱云端算力依赖以控制成本，以透明展示置信度与上下文以重塑操作员信任。Wilton 主张，技术人员与管理者应彻底倒转“手握锤子找钉子”的惯性，始终“从落水人员的真实处境出发，其余一切由此展开”。 Wilton 指出，该研究当前存在明确的边界条件与局限性：模型尚未在恶劣风浪海况、浓雾与海雾等严重气象遮蔽条件下，或多类传感器与机型载荷下完成全要素测试，不能直接作为即用型方案取代现有搜救演练规程。未来迈向实装部署的技术路线图包含三大核心里程碑：在加固适航的机载边缘硬件上推进实测验证、在不同水温盐度与全天候光照条件下扩展数据集代表性，以及将红外视频流与雷达、GPS、船舶自动识别系统（AIS）等多源传感器进行融合，以形成统一的作战态势图景。

Detailed notes — zh

## Background 该视频由 YouTube 频道 Ai4 于 2026 年 8 月 14 日发布，内容为 Coal Face AI 创始人托马斯·威尔逊（Thomas Wilson）发表的主题演讲。视频简介指出，威尔逊结合其在英国皇家海军服役超 20 年的作战经验与海难搜救人员检测研究，探讨如何将一线真实问题转化为产品假设，在投入巨资前通过“基于产品的研究”识别致命风险，避免盲目追逐技术热点。 ## TL;DR 威尔逊批评了当前科技界手握 AI 锤子四处找钉子、盲目追求“撒点 AI 魔法金粉”的喜鹊综合征，主张以“基于产品的研究”五步法让技术凭实力证明价值。他以海上人员落水搜救为具体切入点，展示了在零下 2 摄氏度的严酷湖水中采集 70 多分钟真实无人机热成像视频，并手动精细标注超 10,000 张图像的验证过程。模型最终在 NVIDIA Jetson Xavier 边缘端实现了 97% 精确率与 96% 召回率，摆脱了对云端连接的依赖。他强调，实用的 AI 绝不仅是优化模型精度，更要在人机交互中解决操作员的信心、成本与信任问题，让技术真正服务于水中等待救援的生命。 ## Atmosphere & style 整场演讲务实、干练且充满一线军旅生涯的冷峻幽默感，没有浮夸的科技布道，而是带着南约克郡煤矿工业传统般的沉稳脚踏实地。威尔逊从漆黑深夜甲板上对坠海的真实恐惧讲起，将生死攸关的残酷统计数据与冰冷刺骨的冬泳试验娓娓道来，形成了一种极具现场说服力的叙事张力。听众在冷静克制的工程分析与直击本质的反思中，能深切感受到一种去伪存真的清醒力量——技术不该高高在上，唯有直面一线最真实的痛苦，AI 才能真正救人于狂澜。 ## Key takeaways - [01:44](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=104s) 企业在引入 AI 时常陷入“拿着昂贵锤子满地找钉子”的误区，未从真实业务痛点出发，导致大量技术投资无法转化为实际业务价值。 - [03:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=185s) 警惕技术狂热与“喜鹊综合征”，企业不应问“我们该如何使用 AI”，而应严谨拷问“该特定能力能否解决现实问题，且能否在重金投入前被证伪和验证”。 - [05:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=320s) “基于产品的研究”（Product-Based Research）核心在于防止犯错——研究的最大价值不是证明自己是对的，而是避免浪费数年时间与巨额资金去证明自己是错的。 - [05:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=320s) 基于产品的研究五步框架：始于一线真实痛点、提出可证伪产品假设、锁定最致命风险假设、优先低成本重点验证、最终以证据驱动后续分阶段投资。 - [06:53](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=413s) 海上搜救面临极高的死亡率（英美水域落水事故致死率超 40%），但海事行业采纳 AI 存在三大核心阻碍：信心缺失、系统集成成本感知过高以及操作员缺乏信任。 - [09:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=545s) 在零下两度的真实公开水域环境中，通过手工标注超 10,000 张无人机热成像图像验证核心假设，揭示了从小样本向实用级模型跨越的关键转折与边际效应递减规律。 - [12:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=740s) 海上极端环境常处于离线状态，必须采用如 NVIDIA Jetson Xavier 等低成本商用边缘计算硬件进行实时推理，模型在验证中达到了 97% 精确率与 96% 召回率。 - [14:25](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=865s) 实用型 AI 的人机交互必须满足清晰度、速度、信任度与上下文感知四大要素，AI 负责边缘预警与航向建议，但决策权必须牢牢掌握在人类指挥官手中。 - [16:08](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=968s) 严谨的研究必须清晰划定未经验证的边界（如恶劣海况与浓雾气象），后续演进需扩充数据集代表性并与航海雷达、GPS、AIS 展开多传感器融合。 - [17:19](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=1039s) 攻克行业落地壁垒的关键在于产品为先：用真实环境实测建立信心，基于商用边缘硬件控制成本，显式呈现模型置信度与判断逻辑以打破黑盒赢得信任。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=4s) 讲者背景与研究初衷：从皇家海军到“煤矿工作面” 来自英国南约克郡的 Coal Face AI 创始人托马斯·威尔逊（Thomas Wilson / 昵称 Tom Wilton）登台开场。他首先解释了公司名称的深厚渊源——南约克郡有着悠久的煤矿开采工业传统，而“Coal Face”（煤矿工作面）在英语中意指最前沿、最切实的实干现场，公司正是致力于在“最需要的地方”交付切实可行的 AI 技术。在创立公司之前，托马斯在英国皇家海军服役了整整 21 年，并在职业生涯晚期任职于皇家海军的“AI 转型小组”（AI Transformation Cell）。二十余年的航海生涯中，他在海上度过了无数岁月，曾与多家搜救机构并肩作战，亲眼目睹了在茫茫大海上搜寻落水者究竟有多么困难。这项研究最初起源于托马斯的硕士学位论文，论文有一个非常学术且冗长的标题——《基于 YOLO 的海难搜救 AI 驱动人员检测》。表面上看，这是一个关于无人机、热成像技术以及利用 YOLO 进行目标检测的纯技术项目。但在推进研究的过程中，托马斯逐渐意识到，技术本身反而是最不值一提的部分，真正引人深思的是“我们为什么要造这个东西”。这个顿悟彻底改变了他对组织机构如何引入 AI 的认知。他明确表示，这场演讲展示的不仅仅是技术工作和实验数据，更重要的是传递一种全新的思维方式，帮助各类机构在未来避免浪费数百万美元的资金与数年的宝贵时间。 - 讲者经历：英国皇家海军服役超 21 年，曾任职于海军 AI 转型小组。 - 企业理念：Coal Face AI 植根于南约克郡煤矿工业遗产，强调在业务一线最务实的地方落地 AI。 - 研究课题：《基于 YOLO 的海难搜救 AI 驱动人员检测》（利用无人机热成像与目标检测算法）。 --- ### [01:44](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=104s) 警惕技术狂热：手握锤子满地找钉子的 AI 陷阱托马斯指出，当前整个行业正沉浸在一种狂热的节奏中：每周都有新的 AI 大模型发布，每月都有大企业宣布全新的 AI 战略，各类科技大会上充斥着成百上千场令人眼花缭乱的现场技术演示。然而，在所有的喧嚣背后，绝大多数企业和机构依然在艰难挣扎，无法从 AI 投资中获得任何切实可见的商业或业务价值。究其根源，是因为大家都问错了问题——大家并不是从亟待解决的“现实问题”出发，而是从“技术本身”倒推。企业先花重金买了一把名为 AI 的昂贵锤子，随后开始四处绝望地寻找可以敲击的钉子。为了扭转这一局面，托马斯提出了“基于产品的研究”（Product-Based Research）这一核心方法论。在此次演讲中，他旨在通过无人机搜救的具体案例，帮助听众建立一套严谨的问题导向型框架，让企业在投入巨资前学会自我审视与验证。无人机与热成像只是今天的实例载体，这套能够识别风险并阻止错误投资的方法论才是真正需要带走的精髓。 ```mermaid flowchart TD accTitle: 传统技术主导与产品主导研究思维对比 subgraph Bad["传统误区：技术优先（昂贵的锤子）"] B1["狂热采购 AI 工具/大模型"] --> B2["四处寻找可用场景（找钉子）"] B2 --> B3["巨额投资难以落地，缺乏实际价值"] end subgraph Good["推荐方案：基于产品的研究（问题导向）"] G1["锚定一线真实业务痛点"] --> G2["提出可证伪的产品假设"] G2 --> G3["识别并验证最致命的风险假设"] G3 --> G4["以真实证据驱动后续分阶段投资"] end ``` --- ### [03:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=185s) 刺破“喜鹊综合征”：把 AI 赶下神坛，让它凭实力争取席位在与各大机构打交道的过程中，托马斯最常听到的问题就是：“我们该如何使用 AI？”（How can we use AI?）。这句话表面上听起来合情合理，但在他看来，这恰恰是企业能提出的“最昂贵、最具破坏性”的问题之一，因为它是典型的“技术优先”思维，默认 AI 必定在企业内部有其容身之所。相比之下，企业真正应该拷问的是两个更朴素的问题：“这项特定的 AI 能力能否解决某个现实中的具体问题？”以及“在我们投入大笔资金之前，能否先证明这一点？”前者是在盲目迎合趋势，而后者则是要求 AI 必须靠实际表现来为自己“赢得席位”。为了展现这种狂热思维的普遍性，托马斯分享了自己在皇家海军 AI 转型小组工作时亲身经历的一桩真实轶事。当时，负责建造英国下一代战略核潜艇的英国潜艇交付局（Submarine Delivery Agency）技术人员前来咨询，对方正在为全新的“无畏级”（Dreadnought-class）战略核潜艇开发操作系统，提出的问题竟然是：“我们能在里面塞进什么 AI？能不能在系统里撒点‘AI 魔法金粉’（Magic AI Dust）？”对方关心的根本不是有什么现存难题可以用 AI 攻克，而仅仅是想在国家最高级别的核威慑装备上强行贴上 AI 的标签。托马斯将这种盲目追逐科技热点的行为比喻为“喜鹊综合征”（Shiny Object Syndrome）——人类天生如喜鹊般被发光发亮的新鲜事物吸引，而 AI 正是当下最耀眼的物件。但企业的失败从来不是因为购买了 AI，而是因为在尚未证明自己是否真正需要它之前，就过早盲目买单。 - 关键反思：从“我们如何用 AI？”转向“这项 AI 能否解决具体痛点？能否在投资前完成验证？”。 - 军方真实案例：英国潜艇交付局在开发无畏级核潜艇操作系统时盲目询问“能撒什么 AI 魔法金粉”，暴露了脱离实际需求的狂热心态。 - 核心论点：“喜鹊综合征”代价高昂；必须让证据来证成技术，而非用技术强推场景。 --- ### [05:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=320s) “基于产品的研究”五步法：研究的价值在于防止犯错针对上述痛点，托马斯正式拆解了“基于产品的研究”五步框架。这套流程拒绝以技术漏洞为由头发起项目，更不接受“因为公司买了软件许可”或“因为 CEO 参加完 AI 峰会大受震撼”这类冲动借口，而是严格遵循以下逻辑推进： 1. **从一线真实业务痛点出发**：始于某个使用者在实际作业中切实经历的痛苦与障碍。 2. **构建可证伪的产品假设**：绝不是泛泛而谈“我们来做个 AI”，而是明确提出“我相信某种具体能力可以解决这个特定痛点”，并且坦然接受这个假设有可能是错误的。 3. **精准识别最致命的假设（Riskiest Assumption）**：每一个 AI 构想都有一个致命的阿喀琉斯之踵。以搜救项目为例，核心风险绝不是无人机能不能飞、热成像相机存不存在，也不是 YOLO 目标检测算法能不能运行，而是极其简单的一点——AI 到底能不能在开阔海域中足够可靠地识别出落水人员？ 4. **优先低成本测试该假设，切忌直接造成品**：在测试核心假设之前，绝不提前投入重金去打造功能完备的系统或购买全套外围设备。 5. **以证据驱动投资决策**：每一次成功的实验为下一阶段的资金投入赚取许可；而每一次失败的实验，都在帮助企业避免浪费数年的研发光阴和数百万经费。托马斯掷地有声地指出：“研究的存在不是为了证明我们是对的，研究的存在是为了防止我们白白浪费数年时间去证明自己是错的。” ```mermaid flowchart LR accTitle: 基于产品的研究五步框架 P1["1. 发现真实痛点 (非技术驱动)"] --> P2["2. 提出产品假设 (承认可能出错)"] P2 --> P3["3. 锁定致命假设 (识别成败核心)"] P3 --> P4["4. 单项重点验证 (暂不造完整产品)"] P4 --> P5["5. 证据驱动投资 (用验证换资金)"] ``` --- ### [06:53](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=413s) 冰冷深海的真实痛点：落水搜救与海事行业的采纳困境随后，托马斯开始将这一框架具体应用到海上落水搜救（Man Overboard）的案例中。他展示了自己在皇家海军服役期间所登上的一艘现代化军舰，并生动讲述了一段带有一丝军旅幽默却无比真实的切身体验。在皇家海军的传统中，舰员在下更（Off watch）或休假期间是被允许在舰上饮酒的，托马斯打趣说自己在他服役过的每一艘军舰上都喝过酒。在深夜时分，他常常一个人站在没有开启任何甲板甲板灯的漆黑上甲板上吹风。望着脚下漆黑冰冷的海水，他曾无数次在脑海中闪过极其强烈的恐惧念头：“如果我现在掉下去，会有人来救我吗？有人看见我掉下去了吗？即使有人看见了，他们能及时赶到我身边吗？” 这绝非杞人忧天。托马斯展示了一组令人震惊的真实统计数据：在 2015 至 2023 年间，仅在英国海域就发生了超过 300 起有记录的落水事件；而在 2023 年单一年份，美国海域记录的落水事件就超过 200 起。更残酷的是，在这些落水事件中，有超过 40% 的事故最终导致人员遇难。在科技高度发达的 2020 年代，如此高的死亡率令人难以接受，这无疑是技术应当大有作为的现实场景。而在调研海事行业对 AI 的接纳度时，托马斯发现行业整体的采纳率依然极低，其阻力主要集中在三个核心要素： - **信心（Confidence）**：操作人员对系统的实际有效性缺乏信心。 - **集成成本（Perceived Cost）**：将 AI 模块整合到原本就极其复杂的船载现有系统中的预估成本过高。 - **信任（Trust）**：操作员如果不信任技术，算法精度再高也毫无用处。在经验丰富的搜救人员眼中，搜救任务中最核心的痛点从来不是海图导航、无线电通信或多资产调度，而是最基础的一步——在茫茫波涛中把人给肉眼找出来。 - 关键数据（英国）：2015–2023 年间，英国水域共报告超 300 起人员落水事件。 - 关键数据（美国）：2023 年单年发生超 200 起人员落水事件。 - 死亡率数据：两地落水事故中均有超过 40% 的比例导致人员死亡。 - 海事 AI 采纳的三大阻碍：信心缺失、系统集成成本感知过高、操作员缺乏信任。 --- ### [09:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=545s) 零下两度的冷水试验：在真实环境中构建信任与数据集明确了“搜救人员难以在广阔复杂的海面上单凭肉眼及时发现落水者”这一痛点后，托马斯团队确立了产品假设：利用无人机搭载热成像镜头并结合 AI 目标检测模型，能够在人眼视觉受限的恶劣海况下，更快速、更可靠地识别出遇险人员。如果这项 AI 算法根本无法稳定捕捉水中的人体热信号，那么后续的一切软硬件开发都毫无意义。因此，验证的核心就是把这个假设扔进严酷的真实环境去测试。为了验证假设，托马斯在距离自己家约 3 英里处的一个由废弃煤矿改建的人工湖——曼弗斯湖（Manvers Lake）展开了实地测试。测试时间选在寒冬 2 月，当时地表环境温度低至零下 2 摄氏度（托马斯风趣地补充道，为了照顾在场的美国听众，他特意在早上 Google 换算成了华氏 28.4 度）。即便水温刺骨，当地依然有一群热衷于公开水域冬泳的志愿者，并欣然同意协助进行试验。团队采集了超过 70 分钟的无人机热成像视频。随后迎来了最耗费心力的环节——人工数据标注。尽管市面上有大量的自动标注软件，但为了确保数据集的高度可靠与可信，托马斯亲自逐帧手动标注了超过 10,000 张图像。数据集中不仅包含水中的单人、携带浮标的人员，还包含了成群的鸭子、天鹅等干扰生物，甚至还标注了湖边栈桥上的冰屋桑拿房。在随后的模型训练探索中，托马斯发现随着标注样本量的增加，模型性能呈现出极具指导意义的变化：从最初仅输入 50 张图像时的极差表现，到提升至 1,000 张图像时的性能质的飞跃，再到 5,000 张图像时的数据演进，充分揭示了小样本向实用级模型跨越的关键转折。 - 试验地点：英国南约克郡曼弗斯湖（由旧煤矿改建的湖泊）。 - 试验环境：2 月冬季实测，地表温度 -2°C（28.4°F），真实公开水域冬泳志愿者配合。 - 数据集规模：超过 70 分钟真实无人机热成像视频，纯手工精细标注超 10,000 张图像（涵盖落水者、浮具、天鹅、野鸭及岸上热源干扰）。 - 验证梯度：从 50 张样本的低精度起步，到 1,000 张样本时实现检测性能的剧烈提升，为后续模型选型奠定了可信基石。 ### [12:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=740s) 边际效应递减与边缘计算的必要性在扩充训练数据集时，模型性能迎来了一次大幅跃升，但当数据集再次翻倍时，性能增长的幅度明显放缓。这种“边际效应递减”在工程实践中是非常宝贵的信号，因为在实际落地中，数据采集和人工标注极其昂贵且耗费心力。明确收益曲线的拐点，就能清楚地知道在什么节点继续投入资金只会换来微不足道的提升。在衡量模型效果时，各项指标都有明确的分工：精确率（Precision）回答的是“当 AI 报警说发现有人时，它到底有多大几率是对的”；召回率（Recall）关注的是“水中所有需要救援的目标中，AI 究竟找到了多少”；F1 分数综合了二者的平衡；而 mAP（平均精度均值）则用于衡量检测框定位的精准度。虽然不需要死记硬背每个指标，但关键在于各项核心数据全部朝着理想方向稳步推进。然而，在海面搜救中，把人识别出来仅仅解决了挑战的一半。系统能否以足够快的响应速度支持现场作战是另一道坎。云计算固然拥有惊人的算力，但一艘孤悬在大西洋中部或北冰洋极端海域的舰船，网络连接往往极不稳定甚至完全处于离线状态。因此，模型推理必须全部在边缘端（Edge）完成。得益于边缘硬件性能的飞速演进，边缘设备如今完全能够提供真正可用的实时推理能力，使得作战级落地产品在当下比几年前要现实得多。在本项研究推进时，英伟达 Jetson Xavier 是团队选用的主力边缘硬件，不仅成本相对可控，而且足以强劲支撑所需的推理计算。 - **核心验证指标**：精确率达到 97%，召回率达到 96%，F1 分数达到 96%。 - **硬件选型**：采用低成本且算力充沛的 NVIDIA Jetson Xavier 边缘计算设备，摆脱远洋对云端连网的依赖。 - **阶段性结论**：这些硬核数字证明核心检测机制完全成立，产品假说成功经受住了检验，值得推进下一阶段投资，而这正是“以产品为导向的研究”（Product-Based Research）最核心的使命。 --- ### [14:25](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=865s) 操作员交互设计与“人员落水”搜救构想证明 AI 算法有效，绝不等于一线人员就会去用它。成功的实用型 AI 核心不仅仅在于精妙的算法，更在于最终操作它们的人。要在高压的救援场景中被采纳，系统必须满足四大维度：清晰度（Clarity，操作员能否一眼看懂 AI 汇报的内容）、速度（Speed，能否在几秒钟而非几分钟内做出决断）、信任度（Trust，操作员是否足够相信建议并采取行动）以及上下文（Context，系统能否清晰解释为什么会判定该处有人）。展示的界面原型并非最终产品，而是一个概念探索：探讨 AI 如何丝滑地融入操作员既有的作战工作流中，同时绝不剥夺操作员的控制权——在生死攸关的行动中，人的主观判断必须始终稳居核心。演讲者半开玩笑地提醒大家，演示图中的比例尺并不严谨，“那个人在现实中可没那么大”。但现实情况是，即便在包括英国皇家海军（Royal Navy）在内的现代化海军中，一旦“人员落水”（Man Overboard）警报拉响，全舰的流程依然惊人地原始：所有人飞奔涌向上层甲板，舰桥人员抓起望远镜，其他人则只能依靠肉眼在苍茫海面上吃力搜寻。这正是现代技术亟待攻克的古老难题。在设想的作战概念（CONOPS）中，一旦触发落水警报，无人机立即自主弹射起飞，按照算法预测的搜索航线前去侦察；一旦识别到落水人员，系统立即向操作员发出预警，并计算出母舰应采取的救援转向航向（Course to steer）。 ```mermaid flowchart TD accTitle: 基于无人机与AI的人员落水搜救作战概念流程 subgraph S1["1. 险情告警"] A["触发人员落水警报 (Man Overboard)"] end subgraph S2["2. 空中自主侦察"] B["无人机快速弹射起飞"] C["算法自主规划搜索航线"] B --> C end subgraph S3["3. 边缘检测与决策支持"] D["机载热成像边缘推理检测"] E["向舰桥操作员告警并提供置信度"] F["推荐救援规避/航向 (Course to Steer)"] D --> E E --> F end subgraph S4["4. 人类最终决断"] G["指挥官核验证据做出最终搜救决断"] end A --> B C --> D F --> G ``` - **人机协同四要素**： - **清晰度 (Clarity)**：呈现直观，杜绝晦涩参数干扰。 - **响应速度 (Speed)**：以秒级决策替代分钟级迟疑。 - **可信赖感 (Trust)**：让一线人员敢于依据线索下达航行指令。 - **情境感知 (Context)**：透明化呈现判定逻辑与环境线索。 - **现状反差**：传统搜救主要依赖全员上甲板裸眼搜寻与望远镜侦查，效率极易受限于恶劣海况与生理疲劳。 --- ### [16:08](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=968s) 研究局限与下一步路线图严谨的研究者必须对自己“尚未证明的内容”保持诚实。该项研究目前并未证明系统在巨浪、海况极恶劣条件下的表现，也尚未验证其在大雾、重霾、海面平流雾（Sea Frets）等极端气象下的鲁棒性，更没有穷尽所有的无人机机型和红外传感器，并且绝不主张立即淘汰或替代现有的搜救安全规程。不夸大边界并不可耻，因为这本就不是初期验证阶段的目标。下一步的研究路线将分为三层推进：首先是边缘硬件实测，此前已经验证了计算可行性，后续将把完整模型烧录到专用的机载边缘设备上进行闭环验证；其次是数据集的扩充与成熟化，重点不在于无脑堆砌图片数量，而在于提升代表性，必须覆盖更广泛的海况等级、复杂气象、海水温度、水体盐度、不同昼夜时段以及不同体型肤色的受试人员；最后是进入实操试验与多传感器融合阶段，将红外识别与舰载雷达、GPS、船舶自动识别系统（AIS）深度交织，使每一个附加传感器都能提供互为印证的线索，以多元信息合力赋能人类操作员。 - **未验证的边界情况**：恶劣海况（Rough seas）、大雾与霾（Fog/Haze）、海面平流雾（Sea frets）、全品类光电吊舱适配性。 - **数据集成熟度维度**：海况等级、气象、水温、盐度、全天候时间段、多样化人体特征。 - **传感器融合目标**：AI 视觉模型 + 航海雷达 (Radar) + 全球定位系统 (GPS) + 船舶自动识别系统 (AIS)。 --- ### [17:19](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=1039s) 攻克落地三大屏障：以产品为先的范式革命前期调研揭示了阻碍 AI 在海上搜救领域落地的三大顽疾：信心（Confidence）、成本（Cost）与信任（Trust）。本研究在设计伊始就针对这三点逐一击破。为了解决“信心”问题，测试直接采用真实无人机在真实海域、真实自然环境下采集的实时热成像视频；如果要让搜救人员相信算法，系统就必须在与他们实战作业极度贴近的环境中自证价值。面对“成本”顾虑，许多组织总误以为上马 AI 就意味着必须采购庞大的 GPU 集群和昂贵的云端基础设施，但本系统自始至终围绕通用的商用边缘硬件构建，毕竟“一套昂贵到无法部署实战的 AI，根本没有解决任何人的实际痛点”。针对最难逾越的“信任”高墙——人们天生抗拒未知的“黑盒系统”——系统在推理输出时特意将置信度指标直接暴露给操作员，使指挥员拿到的不仅是一个冰冷的“是/否”结论，而是能够辅助其独立评估、判断该答案是否可靠的解释性信息。这证明了实用 AI 的突破不仅在于打磨算法架构，更在于扫清横亘在人类面前的使用阻碍。代号为“Vantage”的系统最终验证了一个核心命题：即便在某一专业领域起点数据为零，只要坚守“产品为先”（Product-first）的原则，同样能够构建出可部署在边缘端、具备实战生命力、能够拯救生命的 AI 工具。业界惯常的思维往往是先造出一套强大的技术能力，再去满世界找应用场景；而正确的逻辑恰恰相反——必须“从水中那个落水的人开始，其余一切自然水到渠成”。 - **消除落地三大壁垒**： - **信心 (Confidence)**：拒绝实验室仿真，坚持在真实海况下以无人机实拍热成像验证。 - **成本 (Cost)**：坚决基于低成本商用边缘硬件（Commodity Edge Hardware）设计，降低落地门槛。 - **信任 (Trust)**：打破黑盒，向操作员显式暴露模型置信度与证据链，辅助人工决断。 - **方法论提炼**：提出可复现的“以产品为导向的研究框架”——以真实问题为原点，逐层积累证据链，面向实地作业环境逆向设计，最后再审视该产品是否有必要存在，彻底扭转“拿着锤子找钉子”的行业通病。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=4s) 技术本身反而是最不值一提的部分，真正引人深思的是“我们为什么要造这个东西”。 — 托马斯·威尔逊 > [01:44](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=104s) 企业先花重金买了一把名为 AI 的昂贵锤子，随后开始四处绝望地寻找可以敲击的钉子。 — 托马斯·威尔逊 > [03:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=185s) “我们该如何使用 AI？”……表面上听起来合情合理，但在他看来，这恰恰是企业能提出的“最昂贵、最具破坏性”的问题之一 — 托马斯·威尔逊 > [03:05](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=185s) 我们能在里面塞进什么 AI？能不能在系统里撒点‘AI 魔法金粉’？ — 英国潜艇交付局技术人员 > [05:20](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=320s) 研究的存在不是为了证明我们是对的，研究的存在是为了防止我们白白浪费数年时间去证明自己是错的。 — 托马斯·威尔逊 > [06:53](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=413s) 如果我现在掉下去，会有人来救我吗？有人看见我掉下去了吗？即使有人看见了，他们能及时赶到我身边吗？ — 托马斯·威尔逊 > [06:53](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=413s) 在经验丰富的搜救人员眼中，搜救任务中最核心的痛点从来不是海图导航、无线电通信或多资产调度，而是最基础的一步——在茫茫波涛中把人给肉眼找出来。 — 托马斯·威尔逊 > [14:25](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=865s) 即便在包括英国皇家海军在内的现代化海军中，一旦“人员落水”警报拉响，全舰的流程依然惊人地原始：所有人飞奔涌向上层甲板，舰桥人员抓起望远镜，其他人则只能依靠肉眼在苍茫海面上吃力搜寻。 — 托马斯·威尔逊 > [17:19](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=1039s) 一套昂贵到无法部署实战的 AI，根本没有解决任何人的实际痛点 — 托马斯·威尔逊 > [17:19](https://www.youtube.com/watch?v=N8_cPkXcAiI&t=1039s) 必须“从水中那个落水的人开始，其余一切自然水到渠成”。 — 托马斯·威尔逊 ## Mentioned 托马斯·威尔逊 (Thomas Wilson / Tom Wilton), Coal Face AI, 英国皇家海军 (Royal Navy), 皇家海军 AI 转型小组 (AI Transformation Cell), 英国潜艇交付局 (Submarine Delivery Agency), 无畏级战略核潜艇 (Dreadnought-class submarine), 曼弗斯湖 (Manvers Lake), YOLO, NVIDIA Jetson Xavier, Vantage, 自动识别系统 (AIS), GPS, 雷达 (Radar) ## Video description Thomas Wilson, Founder, Coal Face AI What if the greatest value of AI research is not proving that a product will work, but preventing you from building the wrong one? Drawing on more than 20 years of Royal Navy operational experience, I will use my research into AI-driven person detection for maritime search and rescue to show how a deeply understood real-world problem can be translated into a product hypothesis and tested before major investment is committed. The talk will explore how organisations can identify the riskiest assumptions, choose meaningful measures of success and distinguish a promising model from a genuinely valuable product. Attendees will leave with a practical, problem-led framework for moving from research to prototype, trial and informed AI investment. Join this session to learn how to avoid shiny-object adoption and make evidence, rather than enthusiasm, the basis for your next AI decision. [Source](https://www.youtube.com/watch?v=N8_cPkXcAiI)

Source quotes

[0:04] 托马斯·威尔逊: 技术本身反而是最不值一提的部分，真正引人深思的是“我们为什么要造这个东西”。 [1:44] 托马斯·威尔逊: 企业先花重金买了一把名为 AI 的昂贵锤子，随后开始四处绝望地寻找可以敲击的钉子。 [3:05] 托马斯·威尔逊: “我们该如何使用 AI？”……表面上听起来合情合理，但在他看来，这恰恰是企业能提出的“最昂贵、最具破坏性”的问题之一 [3:05] 英国潜艇交付局技术人员: 我们能在里面塞进什么 AI？能不能在系统里撒点‘AI 魔法金粉’？ [5:20] 托马斯·威尔逊: 研究的存在不是为了证明我们是对的，研究的存在是为了防止我们白白浪费数年时间去证明自己是错的。 [6:53] 托马斯·威尔逊: 如果我现在掉下去，会有人来救我吗？有人看见我掉下去了吗？即使有人看见了，他们能及时赶到我身边吗？ [6:53] 托马斯·威尔逊: 在经验丰富的搜救人员眼中，搜救任务中最核心的痛点从来不是海图导航、无线电通信或多资产调度，而是最基础的一步——在茫茫波涛中把人给肉眼找出来。 [14:25] 托马斯·威尔逊: 即便在包括英国皇家海军在内的现代化海军中，一旦“人员落水”警报拉响，全舰的流程依然惊人地原始：所有人飞奔涌向上层甲板，舰桥人员抓起望远镜，其他人则只能依靠肉眼在苍茫海面上吃力搜寻。 [17:19] 托马斯·威尔逊: 一套昂贵到无法部署实战的 AI，根本没有解决任何人的实际痛点 [17:19] 托马斯·威尔逊: 必须“从水中那个落水的人开始，其余一切自然水到渠成”。

Source quotes

[0:04] Tom Wilton: They've bought the hammer, and now they're desperately looking for nails. [3:05] Tom Wilton: How can we use AI? [3:05] Tom Wilton: Does this specific AI capability solve a real-world problem, and can we demonstrate that before we invest? [3:05] Tom Wilton: What AI can I put in it? What magic AI dust can I sprinkle in there? [3:05] Tom Wilton: AI is today's shiniest object. But organizations don't fail because they bought AI. They fail because they bought AI before proving that they actually needed it. [5:20] Tom Wilton: I believe this capability might solve this specific problem. [6:53] Tom Wilton: If I fell in, is anyone coming? Did anyone see me fall in? And if they did, are they going to reach me in time? [17:17] Tom Wilton: Start with the person in the water and everything else follows.

Original transcript

[0:04] My name's Tom Wilton and I'm the founder of Cold Face AI. [0:08] We're proudly rooted in the industrial heritage of the coal mining in South Yorkshire in the UK [0:13] and we're delivering AI where it matters most at the coal face, hence the name. [0:18] Now, before starting my company, I spent just over 21 years in the Royal Navy [0:23] finishing my career working in the Navy's AI transformation cell. And throughout my career, [0:29] I've spent countless years at sea and worked alongside search and rescue organizations at times [0:36] where I saw firsthand just how hard it was to locate someone in the open ocean. [0:41] Today, I'm going to talk about AI but not in the way you might expect. [0:48] So, my research title AI-driven person detection for maritime search and rescue [0:52] a YOLO bit yes yes yes, a very long title and I even have to read it myself. [0:57] And this started as my master's dissertation and on the surface, it's a project about drones, about [1:03] thermal imaging and about object detection using YOLO but somewhere during the research [1:09] I realized that the technology wasn't actually the most interesting part of what I was discovering. [1:14] The interesting part was why I built it in the first place. And that completely changed how I [1:20] think organizations should approach AI adoption. So, today yes, I'll show you some technical work [1:27] and I'll show you some results. What I really want you to leave with is a way of thinking that [1:32] I believe could save organizations millions of dollars in wasted AI investment and wasted time. [1:39] So, I want you to think of this as the value of product-based research. [1:44] Every week another AI model is released. Every month another company announces an AI strategy. [1:52] Every conference has dozens if not hundreds or thousands of exciting demonstrations and I'm sure [1:57] you're going to see some fantastic demonstrations this week. I know I'm excited to get down to that [2:02] floor and see some of those. Yet, despite all of this, organizations are still struggling to [2:08] get real, tangible value from their AI adoption. Why? Because they're asking the wrong questions. [2:17] Instead of starting with a real problem, they're starting with the technology. [2:22] They've bought the hammer, and now they're desperately looking for nails. So, I'd like you [2:27] to think of this presentation as highlighting the value of product-based research. And this is what [2:33] we're going to cover over the 16 and 1/2 minutes that I've got left on the clock. I know we need [2:38] to keep to the timer there. But, firstly, I'll explain what I believe the biggest trap is that [2:42] organizations are falling into. I'll introduce a framework I call product-based research. [2:48] I'll show you a real example using maritime search and rescue. And we'll look at where that goes. [2:53] And I'll leave you with what I think is the most important lesson from this entire project. [2:58] The drone is simply today's example. The framework is what I really hope you'll take away. [3:05] Whenever organizations talk about AI, I hear the same question over and over again. How can we use [3:12] AI? And that sounds sensible. But, I actually think it's one of the most expensive questions an [3:18] organization can ask because it's technology-first thinking. What I think they should be asking is [3:25] something much simpler. Does this specific AI capability solve a real-world problem? [3:32] And can we demonstrate that before we invest? The first question assumes that AI belongs somewhere [3:39] in an organization. The second makes AI earn its place within that organization. [3:48] And we've all seen this happen. All the different questions, all the different questions [3:53] that are coming out when people are starting to adopt AI. And I want to draw your attention [3:58] to the bottom question, which was a genuine question that I fielded whilst working in the [4:03] Navy's AI transformation cell, which came from the submarine delivery and delivery agency in the UK, [4:09] where they asked, "I'm developing the operating system for the new class of dread submarine, [4:15] which is actually the UK's new nuclear deterrent. What AI can I put in it? What magic AI dust can I [4:22] sprinkle in there? Not what problem have I got to solve maybe AI can use to solve it, [4:30] but actually can we just put some AI in it? You know, and that's the level of thinking that really [4:36] came forward in my mind here about this research methodology. [4:41] And shiny object syndrome, it is expensive. The title says it all. [4:46] The magpie perfectly captures human behavior. We are naturally attracted to shiny objects, [4:53] and AI is today's shiniest object. But organizations don't fail because they bought AI. [5:01] They fail because they bought AI before proving that it that they actually needed it. [5:06] What I'll be talking about today is simply a case study demonstrating one principle. [5:13] Start with the problem. Let the evidence justify the technology, not the other way around. [5:20] So, what does a a better approach actually look like? I call it product-based research. [5:26] Everything starts with a real-world problem, not a technology gap. Not because somebody bought a [5:32] license. Not because your CEO came back from an AI conference excited. You all know who you are. [5:39] A genuine operational problem that somebody experiences. [5:44] Only then do we form a product hypothesis? Not let's build AI, but I believe this capability [5:52] might solve this specific problem. A hypothesis that accepts that it might be wrong. Next, [5:59] identify the riskiest assumption. Every AI product has one. Mine wasn't whether drones could fly. [6:05] Mine wasn't whether thermal cameras existed. It wasn't whether Yolo worked. It was much simpler. [6:11] Can AI actually detect a person in open water reliably enough to be useful? And then don't [6:18] build the finished product. Test that assumption first. And that sounds obvious. But it's amazing [6:23] how often organizations spend millions buying everything before testing one thing that actually [6:29] might matter. And finally, let evidence drive investment. Every successful experiment earns the [6:36] next investment. Every failed experiment saves you from wasting time and money. Research shouldn't [6:41] exist to prove that we're right. Research should exist to stop us wasting years of being wrong. [6:50] So let's apply the framework. [6:53] This is me. I told you I was in the Royal Navy, and this is a warship that I served upon. [6:57] Now, 21 years in the Royal Navy taught me a lot of things. I literally spent years at sea. [7:04] Now, the good grace of the Royal Navy allows you to drink alcohol at sea. Now, you can't do it [7:10] whilst you're on watch, but you can actually drink alcohol when you're off watch or you're taking [7:14] some downtime. And I have to say I have drank on every single warship that I've ever served upon. [7:20] And I've also found myself on the upper deck at all hours of the night when it's very dark on a [7:25] warship. We don't have deck lights. And I wondered to myself, if I fell in, is anyone coming? [7:34] Did anyone see me fall in? And if they did, are they going to reach me in time? A genuine problem. [7:43] And unfortunately, this is not a hypothetical problem. Between 2015 and 2023, there were over [7:50] 300 reported man overboard incidents in UK waters. And in the US, in 2023 alone, [7:58] over 200 man overboard incidences. And that figure in the middle, over 40% of those incidents [8:05] resulted in loss of life. In the 2020s. Surely, technology can get after this problem. [8:16] As part of this project, I wanted to understand why the AI adoption in the maritime industry. [8:21] And I can tell you that it is still quite low for a number of reasons. [8:25] Confidence. Confidence in the systems. The perceived cost of integrating those AI systems [8:32] into otherwise complicated on-board systems. And trust. If operators don't trust the technology, [8:39] it doesn't matter how accurate it is. And I want you to remember those three [8:43] as we go through this presentation. Search and rescue experience was high. [8:50] And the biggest challenge that people were saying was finding people in the water. [8:55] Not navigating, not communicating, not coordinating assets. It was actually [9:00] the ability to find somebody in the water. [9:05] So, let's use the methodology. Remember, we're not trying to prove that AI is amazing. [9:10] We're trying to decide whether this product or solution deserves to exist. [9:16] So, point one, our real-world problem was simple. [9:19] Operators struggle to locate a person who has fallen overboard. Not because they aren't skilled, [9:25] but because the sea is enormous, conditions are difficult, and the human eye has limits. [9:31] So, our product hypothesis became AI could act as a scalable enhancement by identifying [9:38] people using thermal drone footage. And next came the most important question, what could completely [9:44] destroy this idea? And quite simply it was if the AI system com- completely missed people, [9:50] nothing else mattered. So, we test that assumption first. Using a real drone, real thermal imaging, [9:58] real water, real people, and real environmental conditions. And only then, after collecting the [10:04] evidence, is it possible to decide whether or not that this product has a place. [10:11] So, let's have a look at what we actually built. The product hypothesis was simple. [10:15] A drone-mounted thermal camera combined with an AI object detection model can reliably identify a [10:20] person in distress at sea faster and in conditions where human visual detection is unreliable. [10:27] It's a hypothesis, it can be wrong. It can be disproved, and good research doesn't try to prove [10:32] itself right. It gives itself every opportunity to be wrong. If this experiment failed, [10:39] that would still have been valuable research because we'd know not to spend the next 5 years [10:44] building the wrong product. So, the trial location took place around 3 miles from my home at a place [10:51] called Mambo's Lake, which was reclaimed as a old colliery site. And even in February in minus [10:58] two surface condi- minus two degrees centigrade surface conditions, that's 28.4° Fahrenheit for [11:04] the Americans in the room. I do apologize for that, but I googled that this morning for you. [11:08] People still enjoyed doing open water swimming, and they agreed to be part of the trial. [11:13] So, throughout this trial was I collected over 70 minutes of thermal imaging drone footage. [11:19] And then came the glamorous part, which was labeling it all manually. [11:24] Now, I'm sure you could say, "Isn't there an app for that?" There absolutely is an app for [11:29] labeling your images, but what I wanted was a validated and trustworthy data set. [11:34] So, I labeled by hand over 70 minutes of footage creating over 10,000 images. [11:42] And these were the This was the data set. You know, we had we had humans, we had humans with [11:46] float. We then had ducks, a lot of ducks, and a lot of swans that all needed manually label. And [11:52] we even got in the the igloo sauna as well on the on the jetty which some people quite liked to use. [12:03] Now, I'm not going to labor on this slide too much, but this graph tells one of the most useful [12:08] stories from the entire project. And I started with just 50 images and performance was poor, [12:13] quite understandable. At 1,000 images, everything improved dramatically as expected. 5,000 images, [12:20] another significant jump. But, then doubling the data set again didn't double the performance. [12:26] The improvements became much smaller, and that's what we call diminishing returns. [12:31] From an engineering perspective, that's fantastic information. [12:35] Because data collection and annotation is expensive and mine them in. [12:41] It tells us roughly where additional investment starts giving progressively smaller benefits. [12:46] And you'll notice several different metrics here. Precision asks, "When the AI says I [12:52] found somebody, how often is it right?" Recall asks, "Out of everyone actually in the water, [12:59] how many did it find?" And the F1 score combines both of those. The mean average precision scores [13:06] measure how accurately the AI places its detection boxes. You don't need to remember every metric. [13:12] The important point is every one of them moved in the right direction. Now, [13:18] finding people is only half of the challenge. Can the system actually run quick enough to be useful? [13:24] Cloud computing is incredibly powerful, but a ship in the middle of the Atlantic Ocean or in the [13:29] Arctic Ocean may not have reliable connectivity. So inference needs to happen at the edge. [13:35] Edge devices are now capable of delivering genuinely useful real-time performance. [13:41] That makes an operational product more realistic than it would have been only a few years ago. [13:46] And at the time of the research, the Jetson Xavier was the edge device of choice. [13:51] Relatively low cost and capable of running the necessary inference. [13:57] So did the hypothesis survive? Yes. 97% precision, 96% recall, 96% F1 score. Those numbers gave me [14:06] the confidence, not that I'd built a finished product. We'll see that from the UI and UX work [14:11] in the next slides. But that the core detection mechanism was credible and this idea deserves the [14:17] next stage of investment. And that's exactly what product-based research is supposed to achieve. [14:25] So let's imagine we've proved the AI works. That still doesn't mean that anyone will use it. [14:31] Because successful operational AI isn't just about algorithms. It's about the people and [14:37] the operators that will use them. So clarity. Can someone understand what the AI is telling them? [14:44] Speed. Can they make decision in seconds rather than minutes? [14:48] Trust. Do they believe the recommendation enough to act upon it? [14:52] And context. Can they understand why the AI thinks there's a person there? [14:59] And this isn't the finished product. It's simply a concept. What I wanted to explore [15:05] was how AI could integrate into an operator's workflow without taking control away from them. [15:11] Human judgment must remain at the center of these types of operation. [15:17] And the next steps. Now, this is not to scale. I will say that now. That person is not that big. [15:24] But, even in today's modern navies, the Royal Navy being one of them, [15:29] when the man overboard alarm goes off, everybody rushes onto the upper deck. The bridge crews, they [15:35] grab their binoculars. And everybody else uses the naked eye to search for that person. Surely, [15:42] technology can get after that. So, as a concept of operations, maybe the man overboard alert happens. [15:50] And then the drone launches. And it can predict its own pathway to go and find and search for [15:54] that person. And then recognize and alert the operator where that person is, so then they can [15:59] recommend a course to steer to find them. That is how technology can get after an age-old problem. [16:08] Good researchers should be honest about what they haven't proven. [16:12] So, this research didn't prove performance in rough seas. It didn't prove it in every weather [16:17] condition, in fog, in haze, in sea frets. It didn't prove every camera and every drone. [16:23] And it certainly didn't prove that this should immediately replace existing procedures. [16:28] And that's okay. Because that's not what this project was trying to achieve. [16:36] But, the research doesn't end here. The first stage is edge hardware testing. [16:40] And so far, I've demonstrated that that is computationally feasible. The next step, [16:45] taking the model, validating validating it on a dedicated edge device. Secondly, [16:51] the data set needs to grow and it needs to mature. Not simply by adding more images, but making sure [16:56] that the data set is representative. So, we need different sea states, different weather, [17:01] different water temperatures, different salinity, different times of day, and different people. [17:06] Then we go on to operational trials and sensor fusion, combining the system with radar, GPS, AIS, [17:12] where each additional sensor provides another piece of evidence. Multiple sources of information [17:17] coming together to support the human operator. So earlier I mentioned that my survey survey [17:24] identified three recurring barriers preventing AI adoption in maritime search and rescue, [17:29] confidence, cost, and trust. I wanted this project to deliberately tackle each one. [17:35] I wanted to use real-time thermal footage captured from live drones over real water [17:40] under real environmental conditions. If operators are going to believe in AI, it needs to prove [17:45] itself in environments that resemble the ones that they actually work in. Many organizations [17:52] assume that AI requires enormous GPU clusters and expensive cloud infrastructure. [17:57] It doesn't have to. From the beginning, I designed this concept around commodity edge hardware. [18:02] The goal was to make it deployable because an AI system that's too expensive to field isn't really [18:07] solving anyone's problems. And finally, trust. This is perhaps the hardest challenge of all [18:13] because people don't trust black boxes. So the inference was deliberately designed [18:18] to expose the model's confidence so the operator doesn't just receive an answer. [18:22] They receive information that helps them judge whether they should trust that answer. [18:27] When you look across all three, you begin to see that successful AI isn't just about models. [18:33] It's about solving the practical barriers stopping people from adopting it. [18:39] And the bigger picture, this contains what I hope you'll remember after today. [18:44] Vantage is the name of the system and this prove that you can build operationally viable, [18:48] edge deployable, life-saving AI starting with zero domain data using product-first principles. [18:56] The technical achievements speak for themselves. But the methodological validation is what [19:01] excites me most because I genuinely believe product-based research is a repeatable framework. [19:08] Start with the problem, build the evidence, design for the environment it will be used in, [19:13] then decide whether the product itself deserves to exist. [19:17] Industry's default is to build capability and hunt for the problems. I think we should invert that. [19:22] Start with the person in the water and everything else follows. Thank you.