Drone detection with radar
How a 77 GHz FMCW radar detects and tracks small non-cooperative drones – and where the approach demonstrably stands today. Project working name: Drone Guardian.
Radar-based drone detection means detecting, tracking and assessing flying objects by the radar waves they reflect – independent of radio traffic, light and line of sight. Ozense is developing a 77 GHz FMCW radar with local edge processing for this purpose: it detects target candidates, tracks them over time, assesses them conservatively and raises alerts. It is not a counter-drone effector and not a production-ready product.
01 · The problem
Small non-cooperative drones can survey enterprise sites, disrupt operations or approach sensitive installations. The customary responses – a visual check by site security, a look at existing camera feeds, and a police report if in doubt – all presuppose that someone noticed something.
Each available technical alternative has a hard limit:
- Remote ID only covers drones that actually broadcast their identification.
- Passive RF detection needs an exploitable radio link between drone and pilot. A drone flying an autonomous waypoint mission stays invisible.
- Cameras need visibility, correct pointing and sufficient object size in frame.
- High-performance specialist radars solve the sensing problem but are economically out of reach for many private sites.
An industry survey on drone protection in Germany lists three main reasons why companies take no protective measures at all: cost, lack of expertise, and legal uncertainty. The barrier to adoption is therefore not primarily missing technology.
Can radar detect autonomous drones with no radio link?
Yes – by virtue of the measurement principle. Radar measures the signal reflected by the object and does not depend on the drone transmitting anything. It additionally measures range and radial velocity directly, which is decisive for tracking and prioritisation, and it works independently of visible light.
That is a property of the sensing principle, not a performance promise. Whether an echo becomes a stable, correctly assessed target candidate depends on range, radar cross section, altitude, motion profile and the clutter present at the site.
03 · The Ozense approach
Detection
The radar continuously transmits chirps in the 77 GHz band. Range and Doppler FFTs turn the echoes into a range-velocity map, from which a CFAR detector produces a candidate list. An adaptive clutter filter suppresses static and recurring scene content. The parametrisation aims to preserve slowly moving relevant targets as far as possible.
- Range-Doppler processing
- Adaptive detection threshold
- Adaptive clutter suppression
Tracking
Individual detections are not yet a target. A Kalman-based tracker links them over time, estimates position and motion state, and bridges short dropouts. Only track stability, motion consistency and observation time turn raw detections into a credible candidate.
- 3D tracker with acceleration model
- Track promotion across several frames
- Coasting through short gaps
Track validity
Before any semantic statement comes the question of whether consecutive detections are spatially and temporally consistent with a single target. Moving vegetation, multipath propagation and processing artefacts all generate track-like structures. This stage is kept deliberately separate so clutter is not treated as an equivalent object class – and because it is structurally the most effective lever against false alarms.
- Physical target track vs. artefact
- Ghost-target handling
- Its own measure in the false-alarm funnel
Semantic target assessment
A valid airborne track is assessed as drone, bird or other airborne object, or Unknown. The basis is micro-Doppler features, radar cross section fluctuation, and track and temporal features. Rotating propellers and beating wings do in principle produce different signatures; whether they separate robustly across sites is the central research question.
- Micro-Doppler
- RCS fluctuation
- Track and temporal features
Unknown handling
Unknown is not a malfunction but a deliberate system decision. A classifier that forces a class on thin or contradictory evidence produces confidently wrong decisions – exactly what a control room cannot use. Where evidence is insufficient, the track is reported as uncertain rather than mislabelled.
- Conservative uncertainty handling
- No forced class
- Defined degradation path
Planned optical verification
Target architecture: on a relevant candidate the system slews one of the customer's existing PTZ cameras, provided visibility, range and target geometry allow it. Radar and image information are combined into a conservative alert decision; the event then goes to the control room or VMS.
- Radar-triggered PTZ cueing (R&D goal)
- No continuous video recording intended
- Handover to control room / VMS (R&D goal)
04 · Current development status
| Aspect | Status |
|---|---|
| Radar capture and tracking | Demonstrated under real environmental conditions |
| Range and site robustness | In site-specific validation |
| Drone/bird/environmental clutter | Robust cross-site assessment in validation |
| PTZ cueing and VMS end-to-end | Pilot-phase objective |
| Weather and false-alarm behaviour | To be validated in longer measurement series |
| Active countermeasures | explicitly not part of the system |
Performance figures are established for defined scenarios and real site conditions. The method and development stages are set out on the validation page.
05 · Which sites this is built for
Ozense develops first for a narrow profile: mid-sized civilian high-value sites with their own security or control-room organisation, existing or integrable PTZ/VMS technology, a concrete drone, espionage or disruption risk, and a spatially limited protection need – sites, in other words, for which premium systems are economically or organisationally oversized.
R&D, testing and prototype sites
Direct exposure to espionage and know-how loss, usually clearly delimited sensitive areas, and often private procurement.
Selected sensitive industrial facilities
High-value production and technical installations – initially as point or sector protection, not a full large-scale perimeter.
Selected data centres
Only where there is a drone-specific scenario, exposed plant, and an existing security organisation – not by industry label alone.
Deliberately not addressed first: airports and public authorities with high reference requirements, public spaces and municipalities, large unmanned sites without a response process, and consumer or residential applications.
Does this match your site?
We are looking for people who know the problem from practice – security managers, installers and integrators. An interview with no intention of running a pilot is valuable too.
Request a research conversationFrequently asked questions
What is radar-based drone detection?
Radar-based drone detection identifies aircraft by the radar waves they reflect — not by their radio traffic and not from a camera image. An FMCW radar measures the range and radial velocity of every reflecting object. The hard part is not receiving the echo but answering what is reflecting: a drone, a bird, moving vegetation, or a processing artefact.
How does radar detect a small drone?
The radar continuously transmits frequency-modulated signals (chirps) and evaluates the echoes. Range and Doppler FFTs produce a range-velocity map, from which a CFAR detector extracts candidates; a tracker links these over time into tracks. Only temporal stability turns an echo into a credible target candidate. Small drones are difficult because their radar cross section is very small — considerably smaller than the targets automotive radar hardware was originally designed for.
Can radar detect autonomous drones with no radio link?
Yes, by virtue of the measurement principle. Radar measures the signal reflected by the object and does not depend on the drone transmitting anything. Passive RF detection requires usable radio emissions; an autonomous flight without such emissions cannot be detected through that channel. Radar does not depend on it. Whether that yields a stable, correctly assessed target candidate depends on range, radar cross section, altitude, motion profile and site clutter.
Can radar tell drones apart from birds?
Yes — this is the established approach. Rotating propellers and beating wings produce distinctly different micro-Doppler signatures, and radar cross section fluctuation differs systematically too. Ozense evaluates these features together with track and temporal features. The real work lies in keeping that discrimination stable across different sites and clutter conditions, which is exactly what the ongoing field validation addresses.
Can Ozense defeat or neutralise drones?
No. Ozense detects, assesses, documents and alerts. Interception or neutralisation measures — jamming, takeover or kinetic effects — are explicitly not part of the system and would each require separate legal authority. The operator alone decides how to respond to an alert.
All answers: FAQ