UK What video analytics does
Video analytics, also called intelligent video analytics or video content analysis, is software that watches video and works out what is happening in it. It detects objects such as people and vehicles, classifies them, follows them from frame to frame and compares what it sees with rules, then raises alerts, counts events or writes searchable metadata. One operator can then supervise hundreds of cameras, because the software points them to the few that matter.
Older video analytics relied on simple rules such as motion in a zone, which raised false alarms from shadows and weather. AI-based analytics recognise what is actually in the frame, so alerts are more reliable and the system can follow a specific target, count objects or read number plates.
How video analytics works
Most systems follow the same pipeline for every camera stream.
-
//
Decode. The compressed H.264 or H.265 stream from each camera is decoded back into frames, usually by a hardware video decoder.
-
//
Detect and classify. An AI model, typically a convolutional or transformer-based detector, finds objects in each frame and labels them: person, car, lorry, bicycle, bag.
-
//
Track. A tracker links detections across frames and gives each object an ID, so the system knows it is the same person moving, not a new one each frame.
-
//
Apply rules. Tracks are tested against rules: crossing a line, entering a zone, staying too long, moving the wrong way, matching a watch list.
-
//
Act and record. Events trigger alarms, relay outputs or messages to the video management system, and the metadata is stored so footage can be searched later.
Common video analytics functions
Analytics engines bundle a set of standard functions. Most sites use a handful of them, chosen for the risk or the question they need to answer.
| Function | What it does | Typical use | Processing load |
|---|---|---|---|
| Intrusion and line crossing | Alerts when a person or vehicle enters a zone or crosses a line | Perimeters, platforms, restricted areas | Low to moderate |
| People counting | Counts people in and out and estimates occupancy | Retail, stations, venues | Moderate |
| ANPR | Reads number plates and checks them against lists | Car parks, access control, enforcement | Moderate |
| Object tracking | Follows targets across frames and cameras | PTZ control, security, traffic flow | Moderate to high |
| Behaviour and safety | Spots falls, fights, missing PPE, people near vehicles | Factories, warehouses, public spaces | High |
| Forensic search | Finds recorded objects by type, colour or direction | Investigations after an incident | Moderate, runs on stored metadata |
A virtual fence is the simplest of these and often the first one a site turns on.
Where video analytics runs
Analytics can run in three places, and many systems mix them. Smart cameras run lightweight models on board, which suits one or two functions per camera. Edge computers and AI NVRs beside the cameras run heavier models across many streams, using a GPU or AI accelerator. Cloud services run analytics on uploaded video or clips, which suits occasional search and reporting more than live alarms.
Running analytics at the edge cuts latency and bandwidth: only events and metadata leave the site, not every stream. It also keeps video on premises, which helps with data protection. Our edge AI explainer covers the approach in general, and our industrial NVRs include Jetson-based AI recorders that record and analyse every camera they connect to.
What affects accuracy
Analytics are only as good as the video they receive. Camera position and angle, lens choice, resolution on target, lighting, frame rate and compression all matter as much as the model. A person who is ten pixels tall in the far corner of a wide-angle view will be missed by most detectors, and heavy compression smears the detail they need.
Models also need to suit the scene. A detector trained on street scenes may struggle with overhead views, thermal images or unusual vehicles, so many projects fine-tune a model on footage from the site. Tracking quality decides how well counts and paths hold up in crowds, which our object tracking FAQ explains. Our object tracking case study tracks targets on a rugged Jetson computer in vehicles.
Privacy and data protection
Video analytics processes images of people, so in the UK and EU it falls under data protection law, and the Information Commissioner’s Office expects a data protection impact assessment for most surveillance systems. Many analytics functions need no identity at all: counting, line crossing and queue length work on anonymous tracks. Edge processing helps further, because video can stay on site while only events and counts are shared. Facial recognition is the exception, with much stricter legal tests, and most industrial and commercial projects avoid it.
Quick answers on video analytics
Video analytics: common questions
What is intelligent video analytics?
Another name for video analytics. It usually implies AI-based detection and classification, as opposed to simple motion detection.
What are examples of video analytics?
Intrusion alerts, people counting, ANPR, queue monitoring, PPE checks and traffic counting. Each turns video into an event or a number.
Is video analytics the same as a VMS?
No. A video management system records, displays and manages cameras. Video analytics interprets the video, often as a module inside the VMS or on a separate edge computer.
Does video analytics need AI?
Not always, but it helps. Basic motion rules work indoors. AI cuts false alarms outdoors and adds classification, counting and search.
How many cameras can one computer analyse?
It depends on the models and resolution. A compact Jetson system may handle a few streams with heavy models, or dozens with light ones. Sizing starts with a test on real footage.
We build AI NVRs and edge AI computers for video analytics as standard, with the operating system, AI runtime and camera drivers installed, and test them with your cameras and models before they ship.