Video analytics and machine vision both run software on camera images, but for different jobs. Video analytics watches areas through network (IP) or CCTV (closed-circuit TV) cameras: people, vehicles, PPE (personal protective equipment), zones and machine states. Machine vision inspects each part with an industrial camera fixed over the line, with its own light and a trigger. Your ceiling cameras may handle area jobs; defects on moving parts need machine vision.
Example (illustrative): a packaging plant wants an alert when someone walks into a forklift lane, and a check that every carton label is readable. Its dock cameras may cover the first job if they pass the camera check below. The second needs a camera, a light and a reject device at each labeler.
Video analytics vs machine vision, side by side
| Video analytics | Machine vision | |
|---|---|---|
| Camera and lens | Dome or bullet cameras, or analog ones via a recorder, often already installed; wide or varifocal lens | Industrial area-scan or line-scan camera with a fixed lens; global shutter for moving parts1 |
| Mounting and lighting | High on a wall or ceiling; the area’s own light, infrared at night | Fixed at a set distance from the part; its own ring, bar, dome or backlight |
| Frame rate and resolution | Continuous video; pixels spread over a whole area | One triggered image per part, or line-scan rows timed by an encoder; pixels spread over one part |
| What it detects | People, vehicles, PPE, zone entry, counts, machine state, license plates | Surface defects, missing or wrong parts, dimensions, labels, print, codes |
| What limits accuracy | Pixels on the target, viewing angle, obstructions, light that changes over 24 hours | Part presentation, light stability, resolution versus the smallest defect |
| Where it runs | On-site server or edge box, the camera itself, or the cloud | Smart camera or industrial PC at the station, fast enough to decide before the next part |
| What it sends to other systems | Events (camera, time, type, snapshot) to a VMS (video management system), phones, the MES (manufacturing execution system) or a CMMS (maintenance system) | Pass/fail to the PLC (programmable logic controller) to reject or stop; results and images to the MES |
| What drives the cost | Streams to process, model work for your scenes, integrations, any new cameras | Camera, lens, light and mount per station; reject hardware; PLC work; setup per product variant |
The deciding difference is control. Machine vision controls the scene, so the software sees every part the same way. Video analytics takes the scene as it comes, so it suits large, slower events. Once you add a fixed mount, your own light and a trigger, you are building an inspection station, which How AI is Transforming Manufacturing Quality Control covers.
How the video reaches the software
ONVIF and RTSP (Real Time Streaming Protocol) are often confused, but they do different jobs. ONVIF defines the web-service interfaces that software uses to discover cameras on the network, configure them and control them.2
The video does not travel over those web services. ONVIF’s streaming specification sends audio, video and metadata over RTP (Real-time Transport Protocol), with RTSP to set up and control the session.3 An ONVIF client asks the camera for its stream address with the GetStreamUri command, then opens that RTSP address.4 Analytics software needs the RTSP stream; ONVIF makes it easier to find and set up across brands.
ONVIF features come in profiles. Profile S covers video streaming and configuration; it is being deprecated, and March 31, 2027 is the last date for product conformance submissions.5 Profile T covers H.264 and H.265 video, imaging settings, and motion and tampering alarms.6
Profile M covers analytics metadata and events, which a device can also send over MQTT, a messaging protocol used in IoT (Internet of Things) systems.7 For new cameras, ask for Profile T, plus Profile M if the camera should send its own events.
Machine-vision cameras use standards maintained by A3, the Association for Advancing Automation. GigE Vision covers high-speed image transfer from machine-vision cameras over Ethernet,8 and USB3 Vision does the same over USB 3.0. Both use GenICam, a standard way to describe camera features, so software can control any maker’s camera the same way.9
What do you want to detect?
Pick the approach per job, not per plant.
| What you want to detect | Approach | Why |
|---|---|---|
| Hard hat or vest missing in a marked zone | Video analytics | People are large in the frame; seconds of delay are fine |
| Person in a forklift lane or at a robot cell gate | Video analytics, as an alert | A zone rule on a wide view; not a safety-rated guard |
| Headcount at posts or in zones | Video analytics | Counts across a wide area, over time |
| Truck plate at the gate | Video analytics with a lane camera | A wide yard view cannot read plates |
| Machine running, idle or stopped | Controller signal first, camera for the reason | The controller knows the state; a camera shows why |
| Scratch or dent on a machined part at line speed | Machine vision | Small defect, one sharp image per part, PLC reject |
| Missing or crooked label on every carton | Machine vision | Each unit seen in the same pose and light |
| Solder and placement faults on a PCB (printed circuit board) | Machine vision: AOI (automated optical inspection) | Tiny features under controlled light |
Machine state is the job both can do. If a machine has a controller with a data port, such as a PLC or CNC (computer numerical control) unit, read its state there. How to Connect CNC and Legacy Machines to Your ERP covers the protocols. A camera aimed at the machine or its stack light adds the reason: an empty post, a jam, an open guard.
Powered industrial trucks, which include forklifts, were the eighth most frequently cited standard of the U.S. Occupational Safety and Health Administration in fiscal year 2025.10 A camera alert on a forklift lane can back up training, traffic rules and barriers; it does not replace them.
Can you use existing CCTV cameras for production monitoring?
Often, for area jobs. Before buying anything, pull a still and a one-minute clip from each camera at the busiest and darkest hours.
| What to check | How to check it | What fails |
|---|---|---|
| Pixels on the target | Measure in pixels the smallest thing you must see, at the far edge of the zone; ask the software supplier for its model’s minimum | Below that minimum, such as a hard hat a few pixels wide |
| Angle and line of sight | Look at where targets stand in the still | A steep top-down view hides vests and labels; racks or pallets block the zone |
| Light across 24 hours | Compare stills from day, night and with the dock door open | Glare, backlight, dark corners; at night many cameras switch to black-and-white infrared, losing vest color |
| Motion | Step through the clip frame by frame as a person or forklift crosses | The event shows in only one or two frames, or every frame is blurred |
| Stream access | Open the RTSP address in a player such as VLC, with an account the software can use; note codec, resolution and frame rate | Only a vendor app or cloud portal shows the video, or the recorder refuses another connection |
| Fixed view | Check whether the camera pans, tilts or zooms on a schedule | A PTZ (pan-tilt-zoom) camera on patrol moves, so zones drawn on the image no longer match the floor |
| Clock | Compare the camera’s time with the MES server | Events land minutes off and match the wrong order or shift |
A new mount, lens or light often fixes a camera that fails one row. If it fails on pixels and motion for a part check, the job needs machine vision. Check the sensor too: a rolling shutter exposes the image row by row, so fast parts come out skewed. A3 notes that machine vision applications measuring moving objects require global shutters for accuracy.1
Run a pilot before you roll out
A still and a clip rule out cameras that cannot work. Only a pilot shows whether detection works in your light, on your shifts. Write each event as a rule a person could check, such as a person without a vest in the yellow zone for more than 2 seconds. Agree the allowed misses and false alarms per camera and shift before anyone labels a frame.
Then test on unseen footage, and live in shadow mode, where alerts go to a log, not to people. Plan the disk space: GB per camera per day = Mbit/s × 10.8, so an assumed 4 Mbit/s stream fills about 43 GB a day. acty.dev pilots video analytics on the client’s own footage for the same reason.
Privacy and the workforce
A camera that checks vests at a packing table also records who stood there and when. In the EU, guidelines from the European Data Protection Board (EDPB) cover video devices under the GDPR (General Data Protection Regulation). They say surveillance purposes should be documented in writing and specified for every camera in use.11
So PPE checks on a security camera belong in that record. The guidelines also note that an employee at work is in most cases not expecting to be monitored by the employer.11
The EDPB asks for warning signs at about eye level, placed so people see them before they enter the monitored area. A sign should give the purposes, who is responsible, people’s rights and the greatest impacts of the processing, and say where to find the full details.11
On retention, the EDPB says footage kept for purposes such as detecting vandalism should in most cases be erased, ideally automatically, after a few days. The longer you keep it, especially beyond 72 hours, the more you must justify it.11
National law or collective agreements, including works agreements, may set specific rules for employee data under GDPR Article 88.11 So bring in HR and worker representatives before go-live.
Agree what each camera watches and why, what happens after an alert, whether alerts serve coaching only or can lead to discipline, and who may view footage. Machine-vision cameras in shielded stations see parts, not people, so these questions mostly concern video analytics.
Footnotes
-
A3 (Association for Advancing Automation), “Global vs Rolling Shutter: Key Differences”, accessed 2026. https://www.automate.org/glossary/global-vs-rolling-shutter ↩ ↩2
-
ONVIF, “ONVIF Core Specification”, version 26.06, June 2026. https://www.onvif.org/specs/core/ONVIF-Core-Specification.pdf ↩
-
ONVIF, “ONVIF Streaming Specification”, version 26.06, June 2026. https://www.onvif.org/specs/stream/ONVIF-Streaming-Spec.pdf ↩
-
ONVIF, “ONVIF Media2 Service Specification”, version 26.06, June 2026. https://www.onvif.org/specs/srv/media/ONVIF-Media2-Service-Spec.pdf ↩
-
ONVIF, “Profile S”, accessed 2026. https://www.onvif.org/profiles/profile-s/ ↩
-
ONVIF, “Profile T”, accessed 2026. https://www.onvif.org/profiles/profile-t/ ↩
-
ONVIF, “Profile M”, accessed 2026. https://www.onvif.org/profiles/profile-m/ ↩
-
A3, “A3 Officially Releases GigE Vision 3.0, Opening New Possibilities in Machine Vision”, May 2026. https://www.automate.org/vision/news/a3-officially-releases-gige-vision-3-0-opening-new-possibilities-in-machine-vision ↩
-
A3, “What Is USB3 Vision?”, accessed 2026. https://www.automate.org/glossary/usb3-vision ↩
-
U.S. Occupational Safety and Health Administration, “Top 10 Most Frequently Cited Standards”, fiscal year 2025 (Oct. 1, 2024, to Sept. 30, 2025), page updated April 2026. https://www.osha.gov/top10citedstandards ↩
-
European Data Protection Board, “Guidelines 3/2019 on processing of personal data through video devices”, version 2.0, adopted 29 January 2020, paragraphs 15, 37, 48, 113, 114 and 121. https://www.edpb.europa.eu/system/files/documents/files/file1/edpb_guidelines_201903_video_devices_en_0.pdf ↩ ↩2 ↩3 ↩4 ↩5
