Vigiloft home

Research and validation methods

Public-dataset work, false-alert research, and separate sensor studies supporting the current RGB prototype. Operational performance remains to be established through controlled field trials.

Sensor research status

These studies are not part of the live RGB detection path.

Thermal

Offline benchmark

RGB + thermal late fusion on public Anti-UAV-RGBT sequences.

Acoustic

Research baseline

Baseline experiments on public audio data.

mmWave Radar

Research baseline

Baseline experiments on public radar datasets.

Passive RF / Remote ID

Separate prototype

Remote ID edge agent, kept separate from the RGB path.

Sensor Fusion

Research

RGB + thermal late fusion has an offline result. Wider fusion is still in the research program.

Offline public-dataset experiment: RGB + thermal fusion

Public Anti-UAV-RGBT sequence. This is an offline experiment, not a deployed capability.

False-alarm suppression research

Five layers are how the research approaches false alarms. Temporal persistence and human review run in the current prototype. Hard-negative mining, cross-modal confirmation, and contextual rules are the next experiments.

  1. 01

    Hard-Negative / Background Mining

    Planned

    Mine difficult non-drone backgrounds — birds, aircraft, insects, clouds, wires, and site clutter — and test the detector against the scenes that cause false alerts.

  2. 02

    Temporal Persistence

    In the prototype

    An alert requires the target to persist across frames, instead of firing on a single detection.

  3. 03

    Cross-Modal Confirmation

    Planned

    Check whether a second, independent sensor cue agrees before an alert is kept, without giving up detection rate.

  4. 04

    Contextual Plausibility

    Planned

    Score a track against motion and scene rules: speed, apparent size, and whether the path is physically plausible.

  5. 05

    Human-Review Triage

    In the prototype

    The review portal gives a person states to confirm, dismiss, or hold an alert, so workload can be counted in review minutes per shift.

A staged R&D program

The program moves from public-dataset baselines, to a Canadian field dataset, to a comparison of sensor combinations on the same operational metrics.

Phase 1

Single-Modality Baselines

Establish honest per-sensor detection rates, false alarms per hour, and operational metrics on public datasets currently used in this repository (Anti-UAV, DroneRF, MMAUD).

Phase 2

Canadian Field Dataset

Collect site-specific data: drones, birds, FPV targets, day/night/rain/snow, urban and industrial backgrounds, with certified pilots and privacy controls.

Phase 3

Multimodal Fusion & Validation

Compare sensor combinations at a fixed 90% detection rate. Measure false alarms per hour, first-alert time, track continuity, and annualized cost per coverage area.

Public research baselines

Three public datasets anchor the baselines: Anti-UAV for RGB and thermal video, MMAUD for synchronized multimodal recordings, and DroneRF for RF classification.

RGB-T fusion baseline

Anti-UAV

318 RGB-thermal video pairs with bounding-box ground truth across diverse scenes including buildings, clouds, day/night conditions.

Multimodal fusion validation

MMAUD (ICRA 2024)

Time-synchronized stereo cameras, mmWave radar, 4-node audio array, and 3D LiDARs — purpose-built for multimodal anti-UAV research.

RF sensing baseline

DroneRF

RF recordings from three drone models in multiple flight modes with background RF baselines for drone detection and classification.

Canadian regulatory considerations

  • Receive-only RF

    RF work in the program is sensing and Remote ID reception. Test setups are planned against Innovation, Science and Economic Development Canada requirements for the spectrum in use.

  • Transport Canada

    Controlled flight testing will be planned under the applicable Canadian Aviation Regulations and site/airspace requirements.

  • Privacy-by-design

    Field-validation procedures will minimize unnecessary personal data capture and define retention, access, and deletion rules before deployment.

  • Defence relevance

    The research is being prepared for Canadian CUAS experimentation opportunities, including the 2027 IDEaS Urban Sandbox.

Interested in discussing a controlled RGB field test?

Discuss Field Validation