The EHS & safety AI glossary

The workflow that assigns, owns, and tracks corrective actions until they are closed.

A version of Chief built for a specific type of task. Think of it like choosing the right specialist for the job - each agent is configured for a different kind of output.

Computer systems that perform tasks normally requiring human intelligence, such as recognizing patterns, interpreting images, or making predictions.

A notification generated when a system detects a defined condition, such as a worker entering a restricted zone or missing required equipment.

The process of labeling images or video so a machine learning model can learn to recognize objects, behaviors, or conditions. Also called labeling.

Face blurring, timestamp removal, and 3D anonymization that protect worker identity while preserving safety context.

Behavior-Based Safety - a program that observes and reinforces safe behaviors.

A positive safety behaviour captured on camera and shared as a recognition record. Best practices are used to reinforce good behaviour - not to trigger corrective action.

The step in a process that limits overall output, where work backs up because capacity there is lower than elsewhere.

A process for addressing the cause of a problem, then stopping it from happening again or spreading elsewhere.

Intenseye’s AI safety assistant. Use it to ask questions about your safety data, generate reports, draft communications, or build presentations in plain language.

A type of neural network designed to process visual data, widely used in computer vision to recognize objects and patterns in images.

AI that interprets video to detect people, objects, and behaviors in real time.

A value a model assigns to a detection, expressing how certain it is that the result is correct. Thresholds on this score help balance sensitivity and accuracy.

An enclosed or partially enclosed area not designed for continuous occupancy, where limited entry and exit or a hazardous atmosphere can put workers at serious risk.

Crowding of people or vehicles in a shared space at the same time. Congestion signals both a safety risk (more interaction points) and an operational drag (flow slows down).

An ongoing effort to improve processes through many small, incremental changes, often driven by the people who do the work. Kaizen is the Japanese term for this practice.

A graph that plots a process measurement over time against calculated limits, used to tell normal variation apart from a change that needs investigation.

A tracked task created to fix a detected risk, with an owner and a due date.

The time it takes to complete one unit of work from start to finish at a given step or process.

Days Away, Restricted, or Transferred rate. A measure of incidents serious enough to keep a worker away from normal duties, per 100 full-time employees per year.

A visual interface that brings safety data together in one place, so teams can track trends, alerts, and performance at a glance.

The policies and controls that define how data is collected, stored, accessed, and protected across an organization.

A branch of machine learning that uses layered neural networks to learn complex patterns directly from large amounts of data, including images and video.

A product or output that fails to meet its specification or requirement. Also called a nonconformance. Tracking defects shows teams where a process breaks down.

The catalog of safety, operations, and quality use cases Intenseye can detect out of the box.

A configured condition that defines what a computer vision system should watch for, such as a specific unsafe behavior, missing equipment, or access to a hazardous area. Often called a use case.

Any period when equipment or a line is not running, whether planned for maintenance or unplanned due to a fault. Unplanned downtime is a leading driver of lost output.

The legal and moral obligation of an employer to take reasonable steps to protect the health and safety of workers and others affected by its operations.

How long people, vehicles, or assets stay in a given area. High dwell time flags bottlenecks, queuing, or workers lingering in a hazard zone longer than the task should require.

Running AI models on-site, on the floor, instead of in the cloud - for real-time speed and data privacy.

Processing data on or near the device where it is captured rather than sending it to a central server, which reduces delay and supports real-time response.

Environmental, Health & Safety - the function Intenseye serves: protecting people, property, and the environment on the floor.

The inspections and audits module that digitizes checklists and walkthroughs.

Intenseye’s ergonomics detection capability. It uses REBA and RULA assessments to evaluate workers’ body mechanics from camera footage and flag elevated musculoskeletal risk before injuries develop.

Continuous ergonomic risk scoring of working bodies using REBA and RULA.

A real risk the system failed to flag.

A detection flagged as a risk that turned out not to be one.

A structured method for identifying how a process or product could fail, how serious each failure would be, and where prevention effort should go first.

The share of units that move through a process correctly the first time, without rework or scrap. Higher yield points to a healthier process.

The number of injuries over a period relative to hours worked, used to compare safety performance across sites or time periods.

The European Union law governing how personal data is collected, processed, and protected.

An emergency safety intervention triggered when the system detects a pSIF-level event during non-routine work. Hard Stop can halt equipment or processes automatically to prevent a serious injury or fatality.

A source with the potential to cause harm.

A visual overlay that shows where hazards or unsafe behaviours are concentrated across a site or camera view. Heatmaps help identify problem areas that might not be obvious from individual alerts.

A ranked framework for reducing risk, from most to least effective: elimination, substitution, engineering controls, administrative controls, and personal protective equipment.

The stage where a trained model is applied to new data to produce a result, such as identifying a hazard in a live video frame.

SIF trends and analytics that show whether risk is going down across sites.

The connection between two systems so they share data and work together - for example, linking detected events to an incident management or EHS platform.

A step-by-step review of a task that identifies the hazard in each step and defines the controls needed to perform the work safely. Also called a Job Safety Analysis.

A production approach that supplies materials and parts exactly when they are needed, reducing inventory and exposing waste in the process.

A measurable value that shows how well a team or process is performing against a defined target.

A metric that measures outcomes after they happen - such as incident rates or days lost. Lagging indicators tell you what went wrong. See also: Leading Indicator.

The delay between an event occurring and the system responding to it. Low latency is essential for real-time safety intervention.

The number of workdays lost due to workplace injuries over a given period. A lagging indicator used to track the severity of incidents across a site or organisation.

A proactive metric that tracks risk before an incident occurs - such as hazard detection frequency, tasks created, or coaching sessions completed. Intenseye is built around leading indicators: catching problems before they become incidents.

An operating philosophy focused on removing waste and anything that adds cost without adding value for the customer.

The path or zone where a worker could be struck, caught, or crushed by a moving object, released energy, or a falling load.

Lockout/Tagout - procedures that isolate energy sources before maintenance so machines cannot start unexpectedly.

A work-related injury that results in a worker being unable to perform normal duties for at least one full shift or day.

Lost Time Injury Frequency Rate - lost-time injuries per million (or 200,000) hours worked. A lagging indicator used to benchmark injury frequency across sites and against industry averages.

Lost Time Incident Rate - the rate of injuries serious enough to cause missed work.

A method of building systems that improve at a task by learning from data, rather than following rules written by hand for every case.

The trained system that takes an input, such as a video frame, and produces an output, such as a detection or a classification.

An injury to muscles, nerves, tendons, or joints caused by repetitive motion, awkward postures, or overexertion. MSDs develop over time and are the risk ergonomic scoring (REBA/RULA) is designed to catch early.

The average operating time between one equipment failure and the next, used to gauge reliability.

The average time needed to restore equipment to working order after a failure, used to gauge maintainability.

An unplanned event that did not result in injury or damage but easily could have. Near-misses are leading signals - catching them is a chance to fix a hazard before it causes harm.

A machine learning structure loosely modeled on the brain, made of connected layers that learn to map inputs to outputs.

Tasks performed outside standard operating procedures - maintenance, repairs, inspections, or any deviation from normal workflow. Non-routine work carries a disproportionately high share of serious injury risk and is a specific focus of Intenseye’s pSIF and Hard Stop detection.

A computer vision task that locates and identifies specific objects within an image or video, such as forklifts, pedestrians, or protective equipment.

A composite measure of manufacturing productivity that combines availability, performance, and quality into a single percentage.

The U.S. federal agency that sets and enforces workplace safety and health standards.

A formal authorization that controls high-risk activities, defining who can perform the work, under what conditions, and what precautions must be in place.

A design or device that prevents an error from happening or makes it obvious right away. The Japanese term translates roughly as mistake-proofing.

A computer vision technique that identifies the position of a person's body and joints, used to recognize postures and movements linked to risk.

A group of AI scenarios that monitor whether workers are wearing the required protective gear - hard hats, vests, gloves, glasses - in the zones where it is mandatory.

Two measures of model performance. Precision is the share of detections that are correct. Recall is the share of real events the system catches.

An early signal - an unsafe act or condition - that reliably precedes serious incidents.

Maintenance triggered by the actual condition of equipment, using sensors and data to service a machine before it fails.

Maintenance performed on a fixed schedule to reduce the chance of failure, regardless of the machine's current condition.

Building privacy safeguards in from the start - edge processing and no facial recognition.

A statistical measure of how well a process fits within its specification limits. A higher value indicates a process that produces fewer defects.

The category of alerts relating to non-routine work hazards - such as working at heights, crane operations, or machine guarding failures. Process safety alerts are tracked as pSIF events. Distinct from compliance alerts, which cover routine PPE and behavioural rules.

A hazard or event involving non-routine work that has a realistic potential to result in a serious injury or death. pSIF events require immediate corrective action. They are distinct from general high-severity alerts in that they specifically involve non-routine conditions where injury likelihood is elevated.

The number of pSIF alerts per 1,000 eligible monitoring hours. Because facilities have different amounts of camera coverage, density normalises alert counts so you can compare sites fairly. A declining pSIF density means your safety interventions are reducing the likelihood of serious injuries - even as monitoring coverage grows.

An AI-generated forecast of where pSIF risk is likely to increase based on current alert trends. Used in the Insights section to help EHS managers act before risk levels rise, rather than reacting after they already have.

The structured set of use cases, workflows, and monitoring configurations in Intenseye dedicated to preventing serious injuries and fatalities. The pSIF Program dashboard shows pSIF alert trends, scenario breakdowns, facility rankings, and consequence severity.

A moment where a worker on foot and a moving vehicle - forklift, truck, AGV - come close enough to risk a struck-by incident. Intenseye detects PVI events to surface high-conflict zones before a collision happens.

The documented set of processes, roles, and standards an organization uses to consistently meet quality requirements.

The EHS Suite section that captures the why behind an incident - five-whys, fishbone, or whichever method the template enables. Ties root causes back to question answers for audit-trail integrity.

The ability to identify and respond to a condition as it happens, rather than reviewing footage after the fact.

Rapid Entire Body Assessment - a standardized whole-body ergonomic risk score.

A work-related injury or illness that meets the regulator's threshold for recording, for example one requiring medical treatment beyond first aid.

The combination of how likely a hazard is to cause harm and how severe that harm would be.

The process of identifying hazards, evaluating the associated risks, and deciding on controls to reduce those risks to an acceptable level.

The total hours during which non-compliance was detected across all monitored scenarios and facilities. Lower is better. Risk hours rising over time is a signal that unsafe conditions are becoming more frequent or prolonged.

Rapid Upper Limb Assessment - a standardized upper-body ergonomic risk score.

A workplace organization method with five steps - sort, set in order, shine, standardize, and sustain - used to create a clean and efficient environment. Strong 5S also removes many everyday hazards.

The total hours during which operations were in full compliance across all monitored scenarios and facilities. The inverse of Risk Hours - use the ratio of safe to risk hours to understand what proportion of working time is spent in compliant conditions.

A defined method for carrying out a task that removes or reduces the risk of injury, agreed and communicated before the work begins.

A systematic review of safety practices, records, and conditions against a defined standard, used to verify compliance and find gaps.

A 0-100 score measuring how well safety protocols are being followed across your plants. Calculated as the percentage of monitored time your workforce operated within safety rules, based on compliance use cases (PPE, behavior, vehicle, housekeeping).

The shared values, attitudes, and behaviors that shape how safety is prioritized across an organization, from the front line to leadership.

A recorded note of a safe or unsafe behavior or condition, used to build a picture of risk and reinforce safe practice.

A 0-100 score measuring how well safety protocols are being followed for a specific scenario at a specific facility. Scores below 50 appear in red; scores above 95 appear in green. The account-wide safety score is the weighted average across all facilities and scenarios.

Intenseye’s hardware device line. Sentinels power the computer vision capabilities on-site and come in variants for different environments and monitoring needs - including standard cameras, thermal, and solar-powered units.

A motorized dome safety camera in the Sentinel suite.

A depth camera for reach-in and struck-by hazards near machinery.

Industrial edge AI compute that processes every camera on-site, in real time.

A camera that detects temperature and electrical hazards, plus cold-chain excursions.

A measure of how serious incidents are, typically expressed as days lost relative to hours worked.

An incident resulting in severe harm or death. SIF prevention is the primary goal of Intenseye’s detection capabilities. The platform focuses on identifying pSIF events - the precursors - before they become SIF outcomes.

A data-driven approach to reducing defects and variation, aiming for process performance where errors are rare and predictable.

A leading category of workplace injury caused by loss of footing or balance, often linked to housekeeping, floor conditions, or working at height.

The structured set of policies, processes, and responsibilities an organization uses to manage safety and drive continuous improvement.

A documented, step-by-step instruction for carrying out a routine task the same way every time.

The use of statistical methods, often control charts, to monitor a process and act before variation turns into defects.

Two of the most common serious-injury mechanisms: struck-by is contact by a moving object or vehicle; caught-in is being pinned, crushed, or drawn into machinery. Both are core targets for real-time detection.

The pace of production needed to meet customer demand, calculated as available working time divided by required output.

The amount of product or work a process completes in a given period.

A standardised measure of how frequently workplace incidents occur, calculated per hours worked. Used to benchmark safety performance across sites or against industry averages. A lagging indicator.

A short, focused safety briefing delivered to a work crew, usually before a shift or task, to reinforce the hazards and controls at hand.

A management approach that builds quality into every function and level of an organization, with a focus on the customer and continuous improvement.

The labeled examples used to teach a machine learning model. The quality and variety of this data shape how well the model performs in the real world.

Total Recordable Incident Rate - recordable incidents per 200,000 hours worked.

A behavior that deviates from safe procedure - bypassing a guard, skipping PPE, entering a restricted zone.

A hazard present in the environment - a spill, an unguarded machine, a blocked exit.

A specific safety scenario the system is configured to detect - such as hard hat compliance, restricted zone violations, vehicle proximity, or ergonomic risk. Each use case corresponds to a trained detection model running on your cameras.

Explore footage with AI - pathways, dwell, congestion, and cycle-time.

How work actually happens on the floor, versus work-as-imagined in the written procedure.

Facilities that recorded no pSIF alerts in the selected period. A zero-pSIF site is a benchmark - it shows what compliant, well-controlled non-routine work looks like, so high-performing practices can be replicated elsewhere.