Three-dimensional digital noise reduction for cleaner low-light video.
Alerting on a bag or item left unattended.
Operators ignoring alerts because too many are false.
Labelling objects in data to create training examples.
Flagging behaviour that deviates from the norm.
Irreversibly stripping data of identifying information.
Automatic Number Plate Recognition — extracting license-plate text from video.
IndoAI's approach of packaging an AI capability as a programmable, deployable app that runs on the edge camera platform.
A tamper-evident record of who accessed or exported video.
Exposure adjustment so subjects aren't silhouetted by bright backgrounds.
The network capacity, in Mbps, available to carry video.
Interpreting actions such as fighting or running.
Data per second used to encode the stream, trading quality against bandwidth and storage.
The rectangle a detector draws around an object.
A cylindrical, outward-facing camera, common outdoors.
Assigning a category label to an image or detected object.
Convolutional Neural Network — the backbone of most vision models.
The algorithm that compresses (encodes) and decompresses (decodes) video.
The probability a model assigns to a detection being correct.
Permission to process personal data for a stated purpose.
Estimating how many people occupy an area.
The entity that decides why and how personal data is processed.
Collecting and keeping only the data actually needed.
The individual the personal data relates to.
The collection of images or video used to train or evaluate a model.
A camera that switches between colour and IR modes by light level.
Automatic IP-address assignment on a network.
Training a small model to mimic a larger one.
A ceiling-mounted camera in a dome housing, common indoors.
India's Digital Personal Data Protection Act, 2023, governing the processing of personal data, including identifiable video.
Digital Video Recorder — an older recorder for analog cameras over coax.
Time a person spends in a defined zone — a retail KPI.
Running inference on or near the camera instead of in the cloud.
An on-site appliance that runs AI inference on existing camera feeds.
Recording onto a card or disk inside the camera or a local appliance.
A numeric vector representing an image, face, or object for comparison.
A short recording bookmarked around a triggered event.
The harmonic mean of precision and recall.
Locating faces in a frame, without identifying them.
Matching a detected face to a known identity.
Key points on a face used for alignment or analysis.
Automatic switch to backup recording when a device or link fails.
Recognising when a person has fallen.
An alert triggered by something that wasn't a real event.
The angular extent of a scene a camera captures, set by lens focal length and sensor size.
Adapting a pre-trained model to a specific use case or site.
Spotting flames or smoke in video at an early stage.
An ultra-wide 180–360° camera that is de-warped in software.
A lens with a single, non-adjustable focal length.
False Negative Rate — how often a system misses a real event.
The lens property, in millimetres, that determines how wide or zoomed the view is.
The count of people entering a space over time.
Quickly finding events in archived footage by attribute.
False Positive Rate — how often a system raises an unnecessary alarm.
Frames captured per second; affects motion smoothness and bandwidth.
Group of Pictures — the spacing between full keyframes in a compressed stream.
A graphics processor often used to accelerate AI inference.
The verified correct labels a model is measured against.
A widely used video-compression standard.
A newer codec that roughly halves bitrate versus H.264 at similar quality.
A visual map of where movement or dwell concentrates over time.
Checking for hard hats or two-wheeler helmets.
Splitting workloads between edge devices and the cloud.
Checking hand-wash or cleaning routines via video.
A fully self-contained keyframe used as a compression reference point.
Impact-protection code indicating vandal resistance.
The chip (CMOS or CCD) that converts light into the video signal.
Running a trained model on live data to produce predictions.
Alerting when a person enters a restricted zone.
Intersection over Union — an overlap-based measure of detection accuracy.
A network camera that streams video over an IP network rather than coax.
Ingress-protection code (e.g. IP66/67) for dust and water resistance.
Infrared LEDs that let a camera see in darkness, rendered as monochrome.
The delay between an event happening and it appearing or being acted on.
An alert when an object crosses a defined virtual line.
Flagging a person lingering in an area beyond a threshold.
License Plate Recognition — the term used outside India for ANPR.
A measure of light level; low-lux ratings indicate low-light performance.
The high-resolution primary stream from a camera, used for recording.
mean Average Precision — a standard accuracy metric for detectors.
Checking whether a face covering is worn.
One million pixels — a shorthand for camera resolution.
Structured tags — objects, time, zone — attached to video for search.
A simple codec encoding each frame as a separate JPEG; high bandwidth.
A trained neural network that performs a task such as detection.
Multi-Object Tracking — tracking many objects simultaneously.
Basic pixel-change detection that triggers recording or alerts.
A varifocal lens whose zoom and focus are driven by motors for remote adjustment.
Sending one video stream to many receivers efficiently over a network.
Non-Maximum Suppression — removing duplicate overlapping detections.
Neural Processing Unit — a chip specialised for AI inference at the edge.
Network Time Protocol — keeps camera and recorder clocks in sync for accurate timestamps.
Network Video Recorder — a device that records and manages IP-camera streams.
Locating and classifying objects in an image with bounding boxes.
Following a detected object across frames.
Running systems on local hardware rather than in the cloud.
An open format for exchanging trained models between frameworks.
An open standard letting CCTV cameras, recorders, and software interoperate across brands.
When a model memorises training data and fails on new data.
A frame stored only as the difference from previous frames.
Tallying individuals crossing a line or occupying a zone.
Information that can identify an individual, such as a face or plate.
Power over Ethernet — delivering power and data to a camera over one network cable.
Locating body keypoints to infer posture or action.
Detecting whether workers wear required safety equipment.
How many horizontal pixels a camera resolves across one metre at a given distance — the core metric for whether it can detect, recognise, or identify.
Of all alerts raised, the share that were correct.
A fixed, blocked-out region a camera never records.
Removing redundant weights to make a model smaller and faster.
Replacing identifiers with tokens that can be reversed under control.
Pan-Tilt-Zoom camera that can be steered and zoomed remotely.
Shrinking a model (e.g. to INT8) so it runs faster on edge hardware.
Measuring queue length and wait time.
Disk redundancy that protects recordings against drive failure.
Role-Based Access Control — limiting who can view or export footage.
Re-Identification — matching the same person or vehicle across non-overlapping cameras.
Of all real events, the share the system caught.
Blurring faces or plates to protect identity.
A defined area where analytics are applied.
The pixel dimensions of the video, e.g. 1080p, 4MP, or 8MP (4K).
How long recorded video is kept before being overwritten.
Real-Time Messaging Protocol — a streaming protocol used mainly for ingest to servers.
Real-Time Streaming Protocol — how IP cameras and NVRs ship live video.
Labelling an image pixel-by-pixel rather than with boxes.
Querying recorded video in natural language, on-site.
Detection of slips, trips, and falls for safety or claims.
Encoding important regions at higher quality to save storage.
Detecting free and occupied parking bays.
Estimating vehicle speed from video.
Secure Reliable Transport — a low-latency, loss-tolerant streaming protocol.
A label for sensors that keep colour images in very low light.
A fixed, manually assigned network address for a device.
A lower-resolution secondary stream used for live preview or analytics.
Spotting an unauthorised person following through a door.
Tera-Operations Per Second — a rough measure of an accelerator's AI throughput.
The process of teaching a model from labelled data.
A virtual line that triggers an event when crossed.
An eyeball-style camera offering easy aiming and reduced IR glare.
A housing built to survive tampering or impact.
A lens whose focal length (zoom) can be adjusted manually or by motor.
Finding similar items by comparing embeddings.
Distinguishing cars, trucks, two-wheelers, and the like.
Condensing hours of footage into a short reviewable clip.
A drawn boundary monitored for crossings or intrusion.
A transformer-based architecture applied to images.
A virtual LAN used to isolate camera traffic from other network traffic.
Video Management System — software that records, displays, and manages many cameras.
Video Surveillance as a Service — cloud-hosted recording and management.
Wide Dynamic Range — processing that balances very bright and dark areas in one frame.
A browser-native protocol for low-latency live video.
Flagging vehicles moving against the traffic flow.
A family of real-time object-detection models (“You Only Look Once”).