AI-powered counter-drone detection can use radar, radio-frequency (RF) sensing, acoustic arrays, visible cameras and infrared/thermal cameras, with CNN or YOLO-family models often supporting the image-analysis layer. The independent question is not “Does this vendor use AI?” but what detection probability, false-alarm rate, target class, range and test conditions are evidenced for this configuration? The 2025 Malviya et al. review lists techniques and limitations, while the European Commission JRC's 2024 fusion study and 2025 detection/tracking report explain the wider sensor and measurement problem. None of the three publications certifies the performance of an individual commercial supplier.
Original research: what is actually established?
IJNRD (September 2025). Ankita Malviya, Rashmi Singh, Damodar Tiwari, Ankur Taneja and Rajnish Choubey published AI-Powered Drone Detection and Tracking Systems: Advancements and Applications in Border Surveillance and Security — A Review, volume 10, issue 9, pages a368–a379. Its Table 1 qualitatively compares CNN, RNN, support vector machines (SVM), decision trees/random forests and YOLO, including the challenge of detecting small objects. Its Table 2 describes radar, RF, acoustic, IR and visible imaging. The tables contain no controlled, standardised cross-vendor trial data and cannot substantiate a guaranteed detection percentage, universal YOLO ranking or a fielded military performance record. The publisher gives the month and year, but an exact day of original publication was NOT ESTABLISHED from the journal record. Check the original publisher metadata.
European Commission JRC (4 December 2024). Grieco, Amendola and Anderson's Counter-drone systems and data fusion, JRC139587, reviews multi-sensor detection, tracking and identification (DTI), fusion-algorithm challenges and a JRC research proof-of-concept platform. Section 3 is a literature meta-study with limitations including mainly static counter-UAS arrangements; it is not an independently validated performance leaderboard for commercial products. JRC reports different fusion approaches, including early and late fusion, and says comparable testing needs consistent metrics and defined distances.
European Commission JRC (27 February 2025). Poppinga, Anderson and Amendola's Technical developments in counter-drone technology: C-UAS detection, tracking and identification technology, JRC140692, supplies definitions, modality limitations and measurement guidance. Sections 3–4 distinguish sensor reach, track association, fusion and complex environmental dependency. It is a technical annual report, not a product certificate or proof that a specific commercial installation achieved a quoted range.
Detection, localisation, tracking, classification and identification are different outcomes
The JRC 2025 report, section 2, treats DTI as a chain of related tasks. Detection notices a potential object; localisation estimates its position; tracking associates observations through time; classification places it into a category such as likely drone; and identification attempts more specific recognition. A thermal camera detecting a moving speck or RF sensing a possible drone-related transmission does not itself establish the target's location, make/model, intent or legal status. Confirmation and human review may be separate functions.
For market intelligence, keep five outcomes in separate fields and avoid treating a machine-learning bounding box as a validated multi-target tracking or autonomous interdiction system. Detection and mitigation are also different product categories; their regulatory authorisations must be examined independently.
Radar, RF, EO/thermal and acoustic: complementary strengths and blind spots
Based on IJNRD 2025 Table 2 (PDF page 8) and JRC 2025 sections 3.1–3.6, the following is an analytical comparison, not universal supplier-specific range or performance data.
| Sensor modality | Useful observation | Critical qualification to test |
|---|---|---|
| Radar | Motion, range and direction where supported by the radar design; potential multi-target tracking | Small radar cross-section, low altitude, clutter, occlusion, overlapping targets, scan geometry and detection thresholds |
| RF sensing | Radio emissions and control or telemetry characteristics where relevant signals are present | RF-silent or autonomous targets, unrecognised waveforms, spectrum crowding, interference, spoofing and localisation uncertainty |
| Visible electro-optical (EO) | Image features and potential visual confirmation; CNN/YOLO object detection | Small target pixels, field of view versus zoom, low light, fog, motion blur, occlusion and camera/compute latency |
| Infrared / thermal EO | Thermal contrast, including at night where sensor and target conditions permit | Thermal contrast, weather, sensor resolution, environment and target propulsion characteristics; not guaranteed night detection |
| Acoustic | Rotor or motor sound patterns where microphones and signal processing can distinguish them | Wind, traffic, ambient noise, limited effective range, quieter aircraft and false sound matches |
JRC's comparative Table 6 provides generalized indicative modality characteristics, not interchangeable procurement specifications. A sensor described as “long-range” without a target class, detection criterion and conditions is not a reliable comparable fact. RF detection also must not be presumed to work for every autonomous or emissions-controlled airframe.
What CNN, RNN, SVM and YOLO actually do
CNN denotes a family of neural-network approaches used to analyse spatial image features; YOLO refers to a family of object-detection architectures that can place bounding boxes and classes on images. The 2025 review's Table 1 also discusses recurrent models for time sequences and classical classifiers for features such as RF data. This is a conceptual technology inventory: the review does not train identical detector versions on a shared public field dataset and does not establish that YOLO is inherently superior or inferior to every CNN implementation.
The algorithm name is not a measured deployment capability. Buyers should request the exact model release, inference hardware, resolution, data domain, target pixel size, confidence/overlap thresholds, test set and held-out-site validation. For tiny drones against birds, aircraft, clouds and foliage, request class-specific precision and recall and confusion examples at the intended range. Training-set “accuracy” and frames per second on a laboratory device must not be promoted into site-wide probability of detection or response speed.
Why fusion may help—and why more sensors do not guarantee fewer errors
JRC 2024, section 3, distinguishes early fusion (closer to raw sensor features), late fusion (combining sensor decisions or tracks) and classical probabilistic / machine-learning hybrids. Each needs sensor alignment, spatial calibration, timestamp consistency, uncertainty handling and correct associations between observations. Incorrect association can merge two drones into one track or split one into multiple alarms.
JRC 2025, section 4.2, provides an illustrative event-site example: increasing RF traffic, incomplete observations and multiple closely spaced aircraft can make previously manageable fusion difficult. This is not a documented independent field trial, and its example distances/timing are not empirical supplier performance values. JRC explains why sensing updates may arrive with different latency and uncertainty even when radar, RF and optical/IR sensors coexist.
Fusion can improve resilience in some tested situations and make performance worse in others. A defensible comparison reports a single-sensor baseline versus a fused system on the same trial scenarios with matched thresholds, targets and exposure. Never assume an additive coverage range, an automatic reduction in false alarms or fault tolerance simply because a marketing diagram shows multiple sensors.
Buyer test sheet: metrics with explicit denominators
The JRC's 2024 report, section 3.2, highlights detection probability by range and false-alarm probability as candidate comparable indicators; the 2025 report, section 4.4, discusses the need for standardised testing, including the COURAGEOUS work behind CWA 18150:2024. These source reports do not give validated manufacturer-to-manufacturer numbers for the configuration below. UXV Searcher recommends collecting the following fields:
| Evidence field | Comparable measurement definition | Common mistake |
|---|---|---|
| Detection probability (Pd) | Drone opportunities detected / ground-truth drone opportunities, at stated range, geometry, target class and alert threshold | Quoting a percentage without drone-opportunity count or without range |
| False-alarm rate | Non-drone alarms per sensor-hour / site-hour, plus environment, area and alert-reset rules | Confusing false-positive fraction on image frames with alarms per hour of deployed surveillance |
| Precision / recall | TP/(TP+FP) and TP/(TP+FN) with the same unit of analysis, confidence threshold and annotated test set | Comparing different target definitions, datasets or class prevalences as if equivalent |
| Tracking quality | Association correctness, track continuity, split/merge frequency, time-to-track and update latency | Reporting detector boxes as validated multi-drone tracking |
| Operational coverage | Verified azimuth/elevation, blind zones, range and conditions for each installed sensor | Using maximum marketing range as continuous site-wide coverage |
| Evaluation provenance | Vendor/independent tester, dates, test site, model/sensor version, target fleet, negative controls, weather and complete sample sizes | Treating lab results or manufacturer claims as independent acceptance |
Illustrative formulas only: for a specified ground-truth protocol, recall = TP / (TP + FN) and precision = TP / (TP + FP); false alarms/hour is an event-rate with a different denominator. No example values in this buyer worksheet should be read as observed JRC or commercial supplier performance. Require uncertainty intervals where trials are small and test data shared among correlated camera frames.
UXV Searcher supplier scorecard and market interpretation
Track a supplier on separate, source-bound axes: modality and component source; named model and OEM role; third-party integration; documented lab trial; independent site trial; contract/award; delivery/acceptance; installed operational deployment; and maintenance/safety/authorisation evidence. A commercial contract confirms a purchasing relationship, not an independently validated Pd, certification or fleet-wide effectiveness. Mark configurations with missing field tests, denominators or site scope NOT ESTABLISHED, rather than assigning a guessed range, accuracy or maturity grade.
For commercial analysis, differentiate sensing hardware, detection/identification analytics, fusion/command software, installation and lifecycle service revenue. The three sources here provide no defensible 2026 global AI-counter-drone market size, unit forecast, licensed model share, annual shipments or vendor revenue. Developing those would require contract volumes, price, geography, period, ownership and avoiding double-counting software embedded in a complete system.
Explore the related AI detection and interceptor procurement analysis, DOE/NIST AI supplier assurance framework, and edge-perception evidence audit.
Questions readers ask
Can AI reliably detect small drones with a YOLO camera alone?
No single YOLO implementation guarantees small-drone detection. Performance depends on object size in pixels, range, lighting, optics, training data, background clutter, thresholds and test conditions. The 2025 IJNRD comparison is qualitative rather than a vendor accuracy benchmark.
What is sensor fusion in counter-drone detection?
It combines information from different sensing systems, such as radar, RF, acoustic and optical/thermal cameras, to support detection, localisation and tracking. Synchronisation, association errors and shared blind spots can still cause missed targets or false alarms.
How should a buyer compare counter-UAS detection accuracy?
Request detection probability and false-alarm rates at stated ranges and target classes, with test denominators, threshold settings, day/night and weather conditions, ground truth, latency, track continuity and independent evaluation. Headline image accuracy is not a full-system performance figure.
Do JRC research reports certify an operational counter-drone supplier?
No. The JRC's 2024 report reviews fusion literature and describes a research proof of concept, while the 2025 report provides technical guidance and testing context. Neither publication certifies a named commercial system or verifies its deployment performance.
Source register and verification scope
Original literature and government reports support the dated observations; all buyer measurement tables are UXV Searcher's interpretation, not a vendor-specific field trial or a government certification finding.
- Malviya et al., AI-Powered Drone Detection and Tracking Systems — A Review, IJNRD2509042 (PDF) · September 2025, 12-page review; Tables 1–2, not a controlled field benchmark.
- IJNRD original journal metadata · Volume 10, issue 9, pp. a368–a379; exact publication day not independently established.
- European Commission JRC139587 — Counter-drone systems and data fusion (full report) · 2024; sections 2–4 including fusion meta-study and research proof of concept.
- JRC139587 publication record · 4 December 2024; DOI 10.2760/6037951.
- European Commission JRC140692 — C-UAS detection, tracking and identification (full report) · 2025; sections 2–4, Table 6, testing context; no guaranteed vendor performance.
- JRC140692 publication record · 27 February 2025; DOI 10.2760/1517220.