Edge or cloud: where video AI should run

Every video analytics deployment settles one architectural question early: where the model runs. Either the video travels to the compute, or the compute sits beside the video. That choice determines the bandwidth the site needs, how quickly the system reacts, what happens when the internet connection drops and where footage of people ends up.

Start with the arithmetic

A 1080p H.264 stream at good quality runs at roughly 2 to 4 Mbps. H.265 delivers a similar picture at around half the bitrate. At 4 Mbps, one camera produces about 43 GB of video per day. Sixteen cameras need 64 Mbps of sustained upload around the clock and generate close to 700 GB a day.

Most business connections are built the other way around, with generous download and modest upload, and that upload is shared with card terminals, phones and back-office systems. Sending every stream off site means reducing resolution, dropping frames or uploading only fragments triggered by motion. Each of those reduces what a model can see. Small objects, distant faces and fast events are the first details to go.

On site, the same streams travel over the local network, where a gigabit switch carries dozens of cameras at full resolution without strain. The model sees every frame the camera produces.

Response time and resilience

An intrusion alert is most useful in the first few seconds. A unit on the local network analyzes frames as they arrive and can close a relay to sound a siren with no round trip to the internet. When processing happens elsewhere, video has to be encoded, uploaded, queued and analyzed before any response can travel back, and every stage adds delay.

Connections also fail, often at the worst moments: storms, power cuts, damaged cables. A site that analyzes its own video keeps detecting and recording through the outage. Where power or internet is unreliable, a unit with LTE and a battery input keeps alerts flowing as well.

Privacy and data protection

Video of identifiable people is personal data under laws such as GDPR. Regulators, customers and staff ask the same questions about it: where it is stored, who can see it and where it travels. When footage is analyzed on site, the answers are short. The video stays in the building, on hardware the organization controls.

What leaves the site is small and deliberate: an alert with a short clip sent to a named recipient, and figures such as hourly footfall or queue length that contain no images. That matters most in hospitals, schools, hotels and public buildings, where a duty of care applies. It also shortens the review by an IT security team, because there is no continuous outbound video stream to assess.

Sizing the hardware on site

Sizing starts with a count of the cameras that need analysis, plus headroom for the cameras the site will add later. Maiyn hardware covers four sizes.

• Maiyn Edge Mini: up to 8 cameras, for a single shop, café or clinic.
• Maiyn Edge Pro: up to 16 cameras, for a larger store, restaurant or office floor.
• Maiyn Edge Max: up to 32 cameras, with LTE and battery input for sites where power or internet is unreliable.
• Maiyn Sentinel: a rack server handling 64 to 128 or more cameras per node, for hospitals, campuses and shopping centers.

Each unit reads streams from any ONVIF or RTSP camera, NVR or DVR already installed, so the existing recorder and cabling stay in place. For an organization with many locations, each site analyzes its own cameras, and Maiyn VMS brings live view, playback, clips and camera health for every site into one place.