Catalog / Computer Vision / VIS-060

Camera-based bird collision avoidance & shutdown control for wind turbines

Stereo or 360-degree cameras (sometimes with radar) detect, classify and track protected birds such as the red kite and trigger turbine curtailment before a collision, for wind-farm operators and permitting authorities.

Parts needed to run it yourselfPrediction & time seriesVision

Collision risk for protected birds such as the red kite blocks or restricts wind projects in many EU regions. Camera systems (often combined with radar) detect and classify target species in the turbine's vicinity, predict the flight path and command the turbine to slow or stop before the bird reaches the rotor. In Germany such systems count as a recognised protection measure in the nature-conservation framework, which can make permits possible for sites that would otherwise be refused.

Problem & value

  • Today: collision risk is handled by distance rules to nests, fixed seasonal or time-of-day shutdowns (for example during harvest or mowing events) and post-construction carcass searches. Blanket shutdowns cost yield; carcass searches only count dead birds after the fact.
  • AI value: curtail only when a sensitive bird is actually at risk, so protection and energy yield are both kept; a documented, measurable mitigation that helps permitting and repowering.
  • Evidence (mixed, check the source): a peer-reviewed Wyoming study found camera-based automated curtailment (IdentiFlight) reduced eagle fatalities relative to a control facility, but markedly increased the number of curtailments [2]. The vendor cites larger reductions (82% rising to 85%) versus human biomonitoring [3] — treat as manufacturer figures. A German WindForS test-field study reported 90% of detections within 750 m in clear weather and 91% within 600 m in poor visibility, and found low-flying target birds were relatively often missed [4].
  • Policy context: the EU push to accelerate renewables increases pressure to resolve species-protection conflicts quickly; Germany's BNatSchG §45b governs collision-risk assessment and mitigation for wind turbines [5].

Variants

  • Onshore wind, raptors (red kite, white-tailed eagle, others): the core case; one unit per turbine or per cluster.
  • Offshore / coastal: seabirds, migration peaks; radar plus camera, harsher maintenance conditions.
  • Bat-focused curtailment: acoustic and weather-based, not camera — see SPEECH-015.
  • Detection-only monitoring: same hardware used to gather pre-construction or post-construction evidence without actuating the turbine (related to VIS-032).
  • Other infrastructure: airports, solar farms and power lines with the same detection logic.

How it works

  1. cameras (stereo / 360°) [+ radar]
  2. on-site detection of flying objects
  3. species / size classification (target vs. non-target birds, other objects)
  4. multi-view 3D triangulation + tracking
  5. flight-path prediction
  6. risk zone check (distance to rotor, speed, heading, permit rules)
  7. curtailment command to turbine controller (SCADA/PLC)
  8. restart when clear
  9. logging, evidence clips, periodic report to authority
  • Detection and classification: small object detectors and classifiers running on edge GPUs at the mast; the hard part is separating a distant target raptor from other birds, insects and clutter against changing light.
  • 3D tracking: two or more calibrated views give distance and altitude; a motion model predicts whether the bird enters the rotor-swept zone within the turbine's deceleration time.
  • Closed loop: the key difference to passive monitoring is that the system actuates a safety-relevant machine; latency, availability and fail-safe behaviour (what happens if the system is offline) are part of the permit.
  • Human-in-the-loop: operators and ornithologists review logged events, tune risk zones and sensitivity, and validate performance during the trial phase required by the authority.

Data preparation

Needed data: camera footage and radar tracks from the turbine site, turbine SCADA states (rotor speed, curtailment events), site permit parameters (risk zones, target species), and reference observations by ornithologists. It lives on edge recorders, the SCADA/historian, and in consultants' reports and spreadsheets on file shares or e-mail.

Typical gaps: no labelled flight tracks, few examples of rare target raptors, unsynchronised clocks between cameras and SCADA, poor calibration documentation, and scattered permit conditions in PDF scans.

Preparation steps: calibrate and document camera geometry; synchronise time sources; collect and label clips with ornithologists, including negative examples such as insects and other birds (CVAT, Label Studio; see PREP-018, PREP-020); digitise permit conditions into a rule table (PREP-006, PREP-011); define who may access evidence clips and for how long (PREP-025). Data inventory and gap analysis help scoping (PREP-001, PREP-002, PREP-005).

Rough share of effort: medium. With a bought system most preparation is site calibration and the trial-phase validation required by the authority; for own development labelling and validation dominate.

Implementation paths

Which path for whom

  • Enterprise (wind farm operators): A, because permit authorities look at recognised systems with documented detection rates per species and the turbine interface and service contract are part of the product.
  • Developer: C, because open detectors and trackers on HW1 edge GPUs allow own research and integrations, but permit acceptance still needs a trial phase.
  • Public (authorities, conservation bodies): C, because for evaluating and validating systems open, reproducible pipelines and independent test data matter more than a product; sovereign hosting of evidence data is preferred.

Turnkey systems including cameras, edge compute, turbine interface and service contracts. Representative vendors: IdentiFlight (US, camera system with Germany/KNE recognition claimed by the vendor [3]), DTBird (Liquen, Spain, EU), BirdVision (Bioseco, Poland, EU), and radar-based offerings such as Robin Radar (Netherlands, EU) and Swiss Birdradar (Switzerland). Check: recognition by the competent state authority, documented detection rates per species, integration with your turbine OEM, service level and data ownership. Choose A for essentially all operators; the certification and permit track record is the product.

Path B — Build on cloud APIs

Not suited to the control loop (latency, offline safety). Cloud use is sensible only for the back office: reviewing logged clips with a multimodal model, drafting authority reports, trend analysis across farms.

Path C — Self-host (open source)

Research and pilot level: open-vocabulary and YOLO-class detectors, classifiers fine-tuned on bird imagery, OpenCV stereo calibration and multi-object trackers, running on Jetson-class edge devices (HW1) with ONNX/TensorRT. Useful for universities, detection-only studies and evidence gathering; a self-built system will not carry authority recognition without a validation study.

Path D — Custom training

Vendors train on large, site-specific labelled libraries of target-species flights. Operators contribute by supplying labelled events and false-alarm clips. Needs thousands of verified tracks per species across seasons, light and backgrounds; independent validation against human observers is the expensive part.

State of the art (2026)

  • Detection range, species discrimination and flight-path prediction improved with deep learning and stereo geometry, but performance depends on site, weather, background and bird size. Published tests show clear weather ranges of several hundred metres and weaker results for low-flying birds and poor visibility [4].
  • Authorities and bodies such as the Competence Centre on Nature Conservation and the Energy Transition (KNE) review the evidence and publish expectations for validation and operation [1]; independent validation studies and Bavarian state-agency work are cited by the family scope as the reference.
  • Trade-off: sensitive settings protect birds but increase curtailment, so yield loss is the main operating cost [2].
  • Limits: false negatives for small or low-flying targets, false positives from other objects, fog/rain/glare, and dependence on turbine response time. Systems complement, not replace, siting and habitat measures.

Shared resources & combinations

  • Shares object detection, tracking and edge deployment with VIS-010 and VIS-044 (counter-UAS, similar sensor fusion).
  • Natural combination with VIS-032 (species identification, camera traps), SPEECH-015 (bioacoustics, bats), VIS-003 (turbine inspection with drones) and VIS-030 (anonymisation if cameras can see people or properties).

Effort, cost & KPIs

  • Path A: months for permit alignment, installation, calibration and a trial phase; recurring service fees; yield loss from curtailment.
  • Path C/D: a research project over at least one season; validation needs human observer comparison.
  • KPIs: detection rate by species and distance, false alarm rate, curtailment minutes and lost energy yield, system availability, response latency vs. rotor stop time, carcass-search results.

People & change

Involve management for the business case and curtailment losses, the site or asset lead as process owner, IT and OT staff for the SCADA interface, IT security for the closed-loop connection to a safety-relevant machine, legal/compliance and the permit authority early, and the data owners of SCADA and monitoring data.

Works council: no in the normal case, because cameras point at the sky and rotor zone, not at employees. It becomes relevant (depends) if technicians are regularly recorded or performance data is derived; then §87(1) no. 6 BetrVG applies, and §90(1) no. 3 requires information when work processes change, for example new shutdown routines for service teams. Public sector: Personalrat. Data protection: personal data only incidentally; ask the DPO to confirm, a DPIA is usually not needed, mask people in stored clips.

AI literacy: support (AI Act Art. 4) operators and ornithologists with training on detection limits and fail-safe behaviour. Communicate the purpose to neighbours, site staff and conservation groups; run a pilot at one or two turbines with the trial phase the authority requires. Typical concern: lost energy yield from false alarms; address with agreed sensitivity KPIs. Humans validate logged events and own risk-zone changes.

See GUIDE-009, GUIDE-013 and GUIDE-011.

Risks, regulation & compliance

  • EU AI Act: classification depends on use. A system that is a safety component in the management and operation of critical digital or energy infrastructure may fall under Annex III (critical infrastructure); a bird-protection controller that only reduces a turbine's speed is arguably outside this, so check the vendor's conformity position. Not an Art. 5 prohibited practice. Machinery and functional-safety rules also apply to the turbine interface [6].
  • GDPR: primarily not personal data; wide-angle cameras can capture people or neighbouring land, so use masking and limited retention.
  • Nature conservation: permit conditions (BNatSchG §45b, state guidance) define target species, validation and fall-back operation [5]; EU Birds Directive applies.
  • Risks and mitigations: over-trust in detection rates (insist on independent validation), downtime (define fail-safe curtailment), vendor lock-in (data export, open interfaces).

Getting started

  1. Clarify with the permitting authority which systems are accepted and what validation is expected.
  2. Compile site species, nest locations and the avoidance-versus-yield trade-off.
  3. Request tests and references from 2–3 vendors with site-like conditions.
  4. Plan turbine interface, network and fail-safe behaviour with the OEM.
  5. Run a trial phase with observers and tune sensitivity before reducing other restrictions.

Sources

  1. KNE — Minimizing bird collisions with wind turbines (conference proceedings) — https://www.naturschutz-energiewende.de
  2. McClure et al. — Effectiveness of an artificial intelligence-based curtailment system (PLOS ONE) — https://journals.plos.org/plosone/article/file?id=10.1371%2Fjournal.pone.0278754&type=printable
  3. IdentiFlight — Performance (vendor) — https://www.identiflight.com/performance/
  4. Test of IdentiFlight on the WindForS test field (Tethys) — https://tethys.pnnl.gov/publications/test-automatic-bird-detection-system-identiflight-test-field-windfors-context-nature
  5. BNatSchG §45b — https://www.gesetze-im-internet.de/bnatschg_2009/__45b.html
  6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj