Machine Vision
What is it
Machine vision is the use of cameras, lighting, optics and dedicated processing to inspect, measure and classify what is seen, often to produce pass/fail results or condition/severity assessments, typically in industrial environments such as manufacturing lines or rail corridors. Sometimes called Industrial Vision, Machine Vision is systems-engineering oriented: integrating sensors, real-time image processing, I/O and control logic into reliable appliances that run unattended.
Computer vision, by contrast, is a broader field of algorithms and models that interpret visual data, using pixel inspection techniques or, increasingly, using probabilistic machine learning and AI techniques, across many domains from social media to medical imaging. It does not typically include the industrial-grade hardware, timing and control integration of machine-vision systems.
In short, computer vision is the ‘brains’ for understanding images; machine vision adds to that eyes, controlled lighting and integration with sensors and actuators to do work in the physical world.
Machine Vision is often mistaken for Machine Learning. These are two different areas but can overlay. Machine Vision generally relates to the capture of imagery in industrial settings and Machine Learning is the application of Neural networks and what is often referred to as Deep Learning on sets of data. Confusingly Machine Vision systems can also include analysis and interpretation of the data that is captured. Traditionally this has been done with more traditional Computer Vision style techniques such as edge detection or ‘rules based’ algorithms. Increasingly it is being completed with Machine Learning based approaches.
Why it matters
In rail infrastructure, machine vision matters because it enables continuous, consistent and repeatable inspection of assets at scale, reducing reliance on manual patrols and subjective visual checks. Properly engineered systems can detect defects earlier (e.g. missing fastenings, spalling sleepers, encroaching vegetation), allowing condition-based maintenance and reducing the likelihood of service-affecting failures.
Machine vision also supports safety and capacity: by automatically monitoring clearances, overhead line equipment, and platform-train interfaces, infrastructure managers can run more trains with fewer disruptive manual blockages for inspection.
Where it is used
Typical rail applications include:
- Track and S&C inspection: linescan or area-scan cameras on inspection trains or in-cab systems to spot missing clips, broken sleepers, ballast pockets (voiding), and geometry-related symptoms in the visual domain.
- OLE and structure gauging: imaging systems measuring wire position, pantograph interaction, and structure clearances, often fused with LiDAR and positioning systems such as GNSS and inertial.
- Civil and lineside assets: automated condition checks on level crossings, signals, signage, cable routes and vegetation growth from video or still imagery.
- Operational monitoring: platform-train interface cameras, hotbox and wheel impact inspection portals, and trespass/obstruction detection around the right-of-way.
Outside rail, machine vision is widespread in electronics, automotive, food and pharma production for high-speed inspection, measurement and robot guidance.
When: key dates
Industrial machine vision emerged in the late 1970s and 1980s with solid-state cameras and early digital signal processing, initially for electronics and automotive inspection. Adoption accelerated through the 1990s.
In rail, early deployments of video-based inspection and gauging systems started appearing in the 1990s and 2000s on dedicated measurement trains and structure-gauging vehicles, later followed by in-service and cab-based systems. The UK saw significant growth in machine-vision based rail monitoring from the 2010s onward, with multiple firms, including One Big Circle (OBC), deploying forward-facing, underframe and lineside imaging systems integrated into routine inspection regimes.
How it works
A typical rail machine vision system follows this chain:
- Acquisition: Industrial cameras (often multiple viewpoints, sometimes linescan) capture imagery synchronised to speed, distance or position; controlled lighting mitigates shadows, glare and night-time conditions.
- Preprocessing: Edge hardware performs steps such as denoising, stabilisation, rectification and colour or contrast normalisation to make features consistent across runs.
- Feature extraction and inference: Classical methods (e.g. template matching) and/or deep learning models identify objects, defects or conditions of interest such as missing fastenings, cracks, vegetation or encroachments.
- Decision and action: Rules or learned models classify results (pass/fail, severity levels), associate them with precise location, and trigger alerts, work orders or further measurement workflows, often via integration into asset management and maintenance planning systems.
A continual challenge of designing and implementing a computer vision system is assessing what can be achieved on the Edge, and what needs to be completed in a central location. Often Edge devices are constrained in terms of power, space and compute, but they benefit from being closer to the source of data being captured. In contrast, central devices are less constrained on power, space and compute, and much more flexible, but inherently data poor, depending on data transmitted to them from the Edge. The aim is often to complete the computer vision at the Edge, but only once you have proven it in the centre.