A drone searches for a person overboard by probability

A man overboard drone has little value if it merely draws neat circles above the last known position. Its real task is to decide where to look first while wind and current move the casualty and enlarge the uncertainty. A Technical University of Denmark preprint released on 3 August 2026 turns that proposition into a measurable simulation result: its leading strategy averages a success rate above 80%, even when launch is delayed by twenty minutes and the visual detector has only a 30% chance of registering a person during a useful observation.

That is an impressive research result, not a product claim. The work is simulation-based, the aircraft is not a certified leisure-boating system, and several assumptions simplify conditions at sea. This boundary is precisely why the paper deserves an owner’s attention. It identifies where automation may recover precious minutes and where crew practice, personal equipment and recovery arrangements still carry the load.

A search area that drifts rather than stands still

The most persuasive component is not the aircraft but the map beneath it. The researchers use the Leeway model, which describes how a floating object drifts under current and wind. An Extended Kalman Filter combines the initial position with vessel measurements for current and wind speed and direction. It produces a probability ellipse that is recalculated every second and expands as uncertainty accumulates.

The stated benefit is computational efficiency. Rather than propagate a large cloud of particles throughout the mission, the filter estimates the mean position and covariance with a light enough workload to run onboard the UAV. The search need not depend on a remote computing link, a meaningful design feature beyond coastal mobile coverage.

Inputs still govern outputs. The accuracy of the initial mark, alarm delay, current data and wind measurements all influence the ellipse. Across boats with very different helm positions and deck plans—from the cruising arrangements represented amongBavaria yachtsto the broad model mix fromBeneteau—the useful question is not simply where a drone box might fit. It is whether an MOB alert and reliable environmental data can reach the system without hesitation.

Five flight patterns expose a different priority

Five search methods are compared. Zigzag, boustrophedon and spiral patterns pursue geometric coverage. Probability Informed Search and Improved Probability Informed Search, or IPIS, instead prioritise cells where the casualty is more likely to be. IPIS also reduces unproductive travel between search nodes.

The authors ran 1,036,800 simulated experiments across combinations of weather, launch delay, detector performance and search strategy. Their virtual UAV flies at an average speed of 20 metres per second and an altitude of 20 metres, with a 30-minute battery. The mission is stopped when the remaining charge is only sufficient to return to the vessel. The simulated downward-facing camera records at 30 frames per second with a 45-degree field of view.

IPIS produces the strongest average result and remains particularly effective as current rises above 1 m/s, when exploiting the probability distribution pays. Yet the comparison is not a one-way victory. Performance falls for every method as wind increases, and at the maximum tested wind speed of 25 m/s the spiral retains a comparatively higher success rate. Sensor quality, delay and sea state therefore influence which route makes sense.

Spray, glare and a moving deck remain outside the model

An owner evaluating safety technology should concentrate on the assumptions. UAV position is treated as exact. Detection is represented at 30, 50 and 100% probabilities, but the model does not recreate the messy false alarms and misses caused by white water, glare, rain, darkness, clothing and partial visibility. The authors explicitly note that GPS and RTK positioning can deteriorate in adverse weather and that camera performance varies with light and unfamiliar scenes.

The paper also stops short of the hardest piece of marine integration: launching and recovering automatically from a platform that rolls and pitches, with rigging, aerials or a tender inside the flight envelope. The constraints around a sailing deck, illustrated by the varied designs on theDufour Yachts page, will differ from family cruising configurations found acrossJeanneau models. These are market references, not suggestions that either builder took part in the research.

The headline percentage needs equal discipline. Above 80% is a finding inside the defined simulator, not a certified probability of saving a life. The model measures visual detection. It does not prove retrieval, delivery of flotation, or survival until the rescue boat reaches the casualty.

The sensible equipment order starts before launch

The research points toward a more coherent hierarchy for spending. Prevention comes first: suitable jacklines and attachment points, a correctly maintained and worn lifejacket, light, sprayhood and tether where appropriate. The UK Maritime and Coastguard Agency also recommends two independent maritime-radio communication methods and advises that at least one additional crew member carry a Class M VHF MOB device fitted with a light.

Next comes the detection chain: personal alarm, an accessible MOB key, an immediate plotter mark, a lookout who maintains visual contact, and a distress call. Finally, the regular crew must practise a recovery method that works with its actual numbers and physical strength. Finding a person from the air does not solve the problem of lifting them from cold, rough water onto a high freeboard.

The paper’s most credible contribution is therefore a software priority: search the highest-probability water first, and continuously revise the route as the area grows. If sea trials eventually validate positioning, detection, launch and retrieval, an integrated aircraft could become an additional safety layer. One operational lesson is available now. A prompt, accurate last-known position and trustworthy environmental inputs are more valuable than a beautifully flown pattern over the wrong patch of sea.