Crawlers, quadrupeds and drones performing routine visual, thermal and gas-detection inspection rounds that would otherwise require a technician in the field.
Computer-vision inspection & defect detection
Camera- and vision-model-based systems that automatically detect corrosion, leaks, mechanical wear or missed safety steps from continuous video or periodic imagery, replacing purely manual visual inspection rounds.
Short entry. This page summarises our current placement. Full, sourced analyses are published first for the technologies closest to practical use; see the list of full analyses.
How it works
Computer-vision inspection applies trained image models to camera feeds. Detection and segmentation models are taught to find specific things, such as corrosion, a leaking flange, a missing guard or a person without the required protective equipment. Anomaly-detection models take the opposite approach: they learn what a normal scene looks like and flag anything that departs from it.
Images come from fixed process or security cameras, from handheld devices, or from robots and drones. Infrared cameras extend the method to temperature anomalies and, with optical gas imaging, to hydrocarbon leaks that are invisible to the eye. Inference may run in the camera, on an edge server or centrally.
Why it matters
Turns cameras already installed for security or process monitoring into a continuous inspection instrument, catching gradual degradation between scheduled human inspection rounds.
Where it fits
The clearest successes are in discrete manufacturing quality control, where lighting and positioning are controlled. In process plants the promising uses are corrosion screening from drone imagery, leak and spill detection, flare and stack monitoring, and conveyor and belt condition. Any application that analyses images of people needs early involvement from the workforce and from whoever is responsible for privacy.
Why we place it at the Over-Promise stage
Widely marketed by both automation vendors and specialist computer-vision startups since 2024; validated false-positive/false-negative rates in harsh, variable-lighting industrial environments are the main differentiator between vendors. Estimated time to the Proven in Use stage: 2–5 yrs. What the stages mean.
Key risk
Model performance degrades on lighting, weather and camera-angle conditions that differ from the vendor's training data — site-specific validation before relying on it operationally is essential.
Questions to ask before committing
- What are the measured false-positive and missed-detection rates on your own site imagery, in poor light and bad weather?
- Who reviews a detection, and how quickly, before anyone acts on it?
- How is the model retrained when cameras move, equipment is repainted or the plant is modified?
- If people appear in the images, what has been agreed about how that footage is used and retained?
Standards and terms to know
- Segmentation
- Labelling which pixels in an image belong to a defect or object.
- Optical gas imaging
- Using an infrared camera tuned to wavelengths that hydrocarbon gases absorb.
- False positive
- An alarm raised when nothing is wrong, the main driver of lost operator trust.
Related technologies
See this entry on the full maturity map · All Control Devices entries · How placements are made. Stage and time-horizon placements are editorial judgements based on public information as of October 2026, not engineering, safety or investment advice; see the Terms of Use.