# Three Senses Are Better Than One: Multi-Sensor Drone Detection: How Sound, Camera and RF Fusion Works

> Every drone sensor has a bad day. New research shows what happens when sound, sight and radio are combined, and what each one adds when the others fail.

- Published: 19 November 2026
- Author: Ghulam Mujtaba, Ph.D. (https://drosonic.com/about.html)
- Publisher: DroSonic OSINT (https://osint.drosonic.com/), the research arm of DroSonic (https://drosonic.com/)
- Series: Sound and Sight, part 2 of 3
- Previous part: The Ear Points the Eye: Acoustic Cueing: How Microphones Aim Cameras and Lidar at Drones (https://osint.drosonic.com/assessments/acoustic-cueing-cameras-lidar.md)
- Topics: Sensor Fusion, Acoustic Detection, Counter-UAS
- Confidence: moderate
- Web page: https://osint.drosonic.com/assessments/multi-sensor-drone-detection-fusion.html
- Cite as: Mujtaba, G. (2026, November 19). Three Senses Are Better Than One: Multi-Sensor Drone Detection: How Sound, Camera and RF Fusion Works. DroSonic OSINT. https://osint.drosonic.com/assessments/multi-sensor-drone-detection-fusion.html
- Reuse: Text and original diagrams © DroSonic. You may quote this report and share its link with credit to DroSonic OSINT; to republish it in full, reuse our diagrams or use it commercially, contact us first. Photos and third-party figures keep the licences shown in their credits.

## Key questions

**What is multi-sensor fusion in drone detection?**
Combining evidence from different kinds of sensor, typically microphones, cameras, radio-frequency receivers and radar, so that one sensor's weakness is covered by another's strength. It can be done by combining each sensor's verdict (decision-level fusion), by combining what the sensors perceive before a verdict (feature-level fusion), or by merging position reports over time in a tracker.

**Does combining sensors improve drone detection?**
In a 2025 Inria study on heavily degraded recordings, the best single sensor (sound, camera or radio) scored 64–67% accuracy, little better than always answering 'drone', and flagged most drone-free clips. Combining all three raised accuracy to 83% and cut false alarms to 3%, though it still missed about one drone clip in four.

**Can sound help a camera detect small drones?**
Yes, in research so far. A 2025 Peraton Labs study fused audio from a smartphone microphone into a camera-based drone detector and improved detection of small, distant drones by 11–15% (mean average precision), with smaller gains for larger drones.

## Key numbers

- **64%** best single sensor, heavily degraded data (about what "always drone" scores)
- **83%** sound, camera and radio combined (false alarms down to 3%)
- **+11–15%** small drones found when sound joins a camera (mean average precision, one study)
- **6 ms** per fused decision on a small edge computer (about 75 millijoules each)

## Bottom line

**No single drone sensor works everywhere, and the research now shows how badly each one can fail on its own.** In a 2025 study from Inria, France's national research institute for digital science, models using sound, camera images or radio signals alone were all close to 97% accurate on clean recordings. On the same recordings deliberately degraded with noise, blur and interference, each fell to 64–67%, about what a system would score by answering "drone" every time, and each flagged most drone-free clips as drones. Combining all three senses raised accuracy to 83% and cut false alarms to 3%. A second 2025 study, from Peraton Labs in the United States, found that adding a microphone's audio to a camera detector improved its detection of small, distant drones by 11–15%. Fusion did not fix everything, since the fused system still missed about one drone clip in four in the worst conditions. But sensors that fail in different ways fail far less often when they are combined well, and acoustic sensing brings something the others lack.

## Which sense fails when

Each sensing method has conditions that blind it. Fusion works because those conditions rarely coincide.

| Condition — what changes | Sound | Camera | Radio receiver |
|---|---|---|---|
| Darkness, glare | Unaffected | Degraded | Unaffected |
| Loud background: traffic, wind | Degraded | Unaffected | Unaffected |
| Crowded radio band: Wi-Fi, other links | Unaffected | Unaffected | Degraded |
| Drone transmits nothing | Unaffected | Unaffected | Blind |

## Every sensor has a bad day

The Inria authors sum up the problem plainly. Acoustic systems struggle in noisy urban settings; camera-based detectors become unreliable in low light or when the drone is partly hidden; radio-frequency detectors, which listen for the link between drone and pilot, are confused in congested areas where other signals overlap, and can be beaten by drones that change frequency ([When Jamming Stops Working](https://osint.drosonic.com/assessments/drone-jamming-limits.html)). ([arXiv](https://arxiv.org/abs/2504.06417)) Radio detection has a further blind spot: drones steered through a fibre-optic cable, or flying autonomously, emit no control signal to find. Our report [When the Detector Stays Silent](https://osint.drosonic.com/assessments/fibre-optic-drones-acoustic-detection.html) covers that case.

Earlier work pointed the same way. A 2021 study from Sweden's Halmstad University combined visible and thermal cameras with a microphone and found that fusion made the system more robust than any single sensor and reduced false detections; the thermal camera did about as well as the visible one despite its lower resolution. ([arXiv](https://arxiv.org/abs/2007.07396)) A December 2025 review of acoustic drone detection in *AIP Advances* lists ten studies combining sound with cameras, radar or radio, and expects future work to concentrate on synchronising data and smarter ways of fusing it. ([AIP Advances](https://doi.org/10.1063/5.0304975)) Fielded systems are moving the same way. A 2024 study of the war in Ukraine by ETH Zurich's Center for Security Studies notes that newer counter-drone systems use AI to analyse signals from radar, cameras, acoustic sensors and video together, while continually updating a threat library of known drone profiles. ([CSS, ETH Zurich](https://doi.org/10.3929/ethz-b-000690448))

## Combine the verdicts, or combine what they perceive

There are two broad approaches. In **decision-level** ("late") fusion, each sensor has its own model that reaches its own verdict, and the verdicts are then combined. It is simple and explainable, and a sensor can be added or removed without retraining the others. In **feature-level** fusion, what each sensor perceives is merged before any verdict, inside a single model. The Inria study used a "gated" version that learns how much to trust each sensor, which can down-weight one that is failing. ([arXiv](https://arxiv.org/abs/2504.06417)) Tracking systems add a third style, merging positions over time, as in the camera-and-microphone system in [Part 1](https://osint.drosonic.com/assessments/acoustic-cueing-cameras-lidar.html).

*Figure: Two ways to combine, verdicts or perceptions.* Diagram of two fusion pipelines. Top, decision-level fusion: microphone, camera and radio receiver each feed their own model, which outputs a verdict such as drone 70%, 40% or 90%; the verdicts are combined into a decision. Note: simple, explainable, add or drop a sensor without retraining. Bottom, feature-level fusion: each sensor's features go into a gate that weighs how much to trust each, then into one model and a decision. Note: learns how the senses relate, can down-weight a failing one. Neither is always better. In the Inria study, the gated approach did best with two sensors; combining verdicts did best with three.

## An open test of sound, sight and radio

The Inria team recorded two small DJI drones, a Mini 2 and a Mini 3 Pro, with an off-the-shelf USB microphone, a compact video camera and a software-defined radio listening on the 2.4 GHz band, in urban and rural places, by day and at sunset. Every sample pairs a quarter-second of sound, seven video frames and one radio snapshot taken at the same moment. They trained on clean recordings, then tested twice: on clean recordings, and on recordings degraded across all three senses at once with added noise, blur, colour shifts and radio jamming, to mimic a bad day in the field. The dataset and code are public. ([arXiv](https://arxiv.org/abs/2504.06417))

On clean recordings almost everything worked: the best single sensors scored 90–98% accuracy and the fused systems about 97%. On the degraded recordings, the single sensors collapsed. The best sound model, the best camera model and the best radio model scored 64%, 64% and 67%. Because 64% of the test clips contained a drone, a model that always answered "drone" would also score 64%, and that is close to what the sound and camera models did: they flagged 100% and 99% of drone-free clips. The best radio model flagged 78%. ([arXiv](https://arxiv.org/abs/2504.06417))

Pairs did better. The best, sound plus radio fused at the feature level, reached 81% accuracy with 11% false alarms. All three senses, combined at the decision level, reached 83% with 3% false alarms. ([arXiv](https://arxiv.org/abs/2504.06417))

*Figure: One, two, three senses, the false alarms are the story.* Bar chart of accuracy and false alarms on heavily degraded recordings. Sound only: 64% accuracy, 100% false alarms. Camera only: 64% and 99%. Radio only: 67% and 78%. Sound plus radio: 81% and 11%. All three: 83% and 3%. A dashed line at 64% marks the accuracy of a model that always answers drone. Alone, each sensor gave up and called almost everything a drone. Together, they rejected almost all the clutter.

The honest caveat is in the same table. In those worst-case conditions the three-sense system caught about three drone clips in four (76%), so it traded misses for far fewer false alarms. Conditions mattered too: under degradation it scored 77% by day, when shifting light and shadow hurt the camera, against 90% at sunset, and 79% in town against 85% in the countryside. Fusion was cheap to run: about 6 milliseconds and 75 millijoules per decision on a small NVIDIA Jetson Orin Nano edge computer. ([arXiv](https://arxiv.org/abs/2504.06417))

> Alone, each sensor gave up and called almost everything a drone. *Together, they rejected almost all the clutter.*

## Small drones, found by ear

The second study asked a narrower question: can sound make a camera-based detector better at drawing a box around a drone in the picture? Peraton Labs recorded a DJI Mavic 3 Pro with nothing more than a smartphone's camera and microphone on a tripod, among trees, buildings, cars and water, by day and in the evening, and built a set of more than 7,500 matched image and one-second audio pairs. They then fed features from a speech-recognition-style audio model into a transformer-based image detector, testing four ways of mixing them. ([arXiv](https://arxiv.org/abs/2509.09859))

*Figure: The method, sound mixed into sight.* Architecture diagram: a camera frame of a drone over trees goes through an image backbone into feature maps at several scales; an audio waveform goes through an audio backbone into audio feature maps; both enter a fusion layer, then a transformer encoder and decoder that output a box around the drone. Audio features join the image features before the detector decides where the drone is.

The gated method again did best. Its biggest gains came where cameras are weakest, on small drones occupying few pixels: detection scores (mean average precision) rose 11.1% on scenes like the training ones and 15.3% on scenes with different backgrounds, weather and light. Across all drone sizes the gain was smaller, 3.3–5.8%. ([arXiv](https://arxiv.org/abs/2509.09859)) That is the complement of Part 1: there, the ear pointed the eye; here, the ear helps the eye recognise what it is already looking at.

## Acoustic panels on the test range

Armies are testing the same idea outside the laboratory. In June 2025, at the U.S. Army's Project Flytrap trials in Hohenfels, Germany, soldiers of the 2nd Cavalry Regiment set up Squarehead acoustic panels alongside other counter-drone systems, and, in the Army's words, tested "the ability of the systems to work both independently and cooperatively with each other." ([U.S. Army](https://commons.wikimedia.org/wiki/File:U_S_Soldiers_continue_C-UAS_testing_during_Project_Flytrap_(9105589).jpg)) The same trials appear in [Watching the Bases](https://osint.drosonic.com/assessments/drones-over-military-bases.html), on drone incursions over military installations.

## What the studies do not show

Both studies are honest about their limits, and a reader should be too. Each covers one or two small DJI quadcopters; neither includes fixed-wing drones, larger military types or swarms. The Inria team's worst-case conditions were simulated by degrading recordings, not observed in the field, and they note the need to add larger drones with different sound and radio signatures. ([arXiv](https://arxiv.org/abs/2504.06417)) Results on research data are usually higher than results in deployment, because field conditions differ from the ones a model was trained and tested on.

For anyone buying or integrating a multi-sensor system, the questions follow:

1. **What does each sensor score on its own, in bad conditions?** A fused figure without single-sensor baselines hides which sensor is doing the work.
2. **What is the false-alarm rate, not just the accuracy?** In the Inria results, the single sensors' accuracy looked middling while their false alarms were near 100%.
3. **Which way are the sensors combined?** Decision-level fusion is easier to audit and extend; feature-level fusion can adapt to a failing sensor but needs retraining.
4. **What happens when one sensor is missing?** A radio-silent drone, a dark night or a loud road should degrade the system, not break it.
5. **Which drones was it tested against?** Two small quadcopters is a start, not a threat library.

### Confidence: How sure we are, and what would change it

**High confidence** that fusing sensors with different failure modes reduces false alarms and missed detections compared with any one sensor. Independent studies from France, Sweden and the United States agree, and a 2025 review describes the field moving that way.

**Moderate confidence** in the size of the gains. The figures come from research datasets of one or two small drones, with degraded conditions simulated rather than recorded, and from a single test set each.

**What would change it:** independent field trials of fused systems against varied drone types, including radio-silent ones, with single-sensor baselines and false-alarm rates published alongside the fused result.

## The sense we build

Of the three senses in this report, DroSonic builds the one that needs neither light nor a radio link: networked passive acoustic sensors that classify drone sound with neural networks on low-cost edge hardware. ([DroSonic](https://drosonic.com/#technology)) It is designed to be fused, not to stand alone; its U.S. provisional patent application covers detection, classification, localisation and tiered defeat cueing. ([DroSonic](https://drosonic.com/news/patent-filed.html))

The studies here show what sound adds when the other senses degrade. DroSonic has not published fused-system results of its own; the prototype tested at DIRT Days 26.04 was evaluated as an acoustic sensor against live drones. ([DroSonic](https://drosonic.com/news/dirt-days-field-test.html))

## The sense that does not depend on light or radio

Fusion is how a defence keeps working when one sensor is having a bad day, which matters more than whether it appears on a datasheet. Of the three senses studied here, sound is the one that needs neither light nor a radio link, which is why it keeps appearing as the complement to cameras and radio receivers, and why Ukraine's acoustic networks, described in [Listening to the Sky](https://osint.drosonic.com/assessments/listening-to-the-sky.html), were built as a layer beneath other sensors rather than a replacement for them.

[DroSonic](https://drosonic.com/) builds that acoustic layer, which is passive, emits nothing an enemy can find, and works in darkness and radio silence. Part 3 of Sound and Sight turns from algorithms to plumbing. It looks at how sensors from different makers share what they detect in practice, through open standards, and where an acoustic layer plugs in.

## Related reports

- [What a Microphone Can Hear: Acoustic Drone Detection Explained: How It Works, Range and Limits](https://osint.drosonic.com/assessments/acoustic-drone-detection-explained.html): Acoustic sensors detect drones by the sound they can't switch off. How they work, how far they hear, and where they beat radar and RF.
- [When the Detector Stays Silent: Fibre-Optic Drones Beat Personal Detectors](https://osint.drosonic.com/assessments/fibre-optic-drones-acoustic-detection.html): RSF's frontline safety guide exposes a gap. Fibre-optic FPVs beat every drone detector a reporter can carry, and only sound is left.
- [Listening to the Sky: Acoustic Drone Detection: WW1 to Sky Fortress](https://osint.drosonic.com/assessments/listening-to-the-sky.html): Ukraine rebuilt a century-old idea, listening for aircraft, with thousands of cheap microphones. How it works, where it fails, and how the U.S. could use it.
