Glossary / Architecture
What is edge AI?
Edge AI means running the model on hardware located where the data is produced instead of sending data to a remote server. For video analytics, a device on site analyses the camera streams and transmits only results. This cuts bandwidth by orders of magnitude, removes network round trips from time-critical decisions, and keeps footage on the premises.
The word edge simply means the far end of the network, the point where data is created. In a video deployment that is the site itself: the factory, the gate, the store, the campus. Edge AI is the decision to do the computing there instead of in a data centre.
This is an architectural choice, not a quality claim. The same model can run in either place and produce the same output. What changes is what has to travel, how long the decision takes, and where the footage lives. Three arguments usually decide it.
The bandwidth arithmetic
Video is heavy and continuous. Analytics results are light and intermittent. That asymmetry is the entire bandwidth argument, and it is easy to quantify.
Worked example: 16 cameras at a single site
32 Mbps × 86,400 s ÷ 8 = 345,600 MB per day ≈ 346 GB per day
≈ 10.4 TB per month leaving the site
Same site, results only:
20,000 detection events × ~1 KB = 20 MB per day
≈ 0.6 GB per month
The ratio is roughly 17,000 to 1. It also changes what uplink you need. Sending the video requires 32 Mbps of committed, sustained upstream, which at most Indian industrial sites is a leased line rather than a broadband connection. Sending the results needs almost nothing, and the video stops being lost whenever the link drops, because the analysis already happened locally.
Note that this cuts the other way at small sites. Two cameras at a retail outlet generate about 1.3 TB a month, which a normal broadband connection handles, and there is no leased line to avoid.
The latency argument
Latency only matters when the output drives an action. A weekly footfall report does not care whether the answer took 40 milliseconds or four seconds. A barrier, a siren, a machine interlock or an operator alert does.
The relevant budget is not the network round trip in isolation. It is capture, encode, transmit, queue, infer, return, act, and every one of those adds. Local inference removes transmit, queue and return from the chain entirely, which typically saves a few hundred milliseconds and, more importantly, removes the variance. A cloud path that averages 300 ms but occasionally spikes to three seconds is harder to design around than a local path that is consistently 80 ms.
Data residency and DPDP
Where video is processed determines where it is stored, who can access it, and how many parties are involved in a breach. Under the Digital Personal Data Protection Act, the operator of the site is generally the entity accountable for the personal data in its footage. Processing on site keeps the number of processors in that chain small and makes the retention question answerable, because the footage never left.
This matters most in the deployments where people are most identifiable: residential gates, employee areas, schools, hospitals. Rule 4 commences on 13 November 2026, with the Section 33 penalty provisions following on 13 May 2027. Reviewing where video is processed before then is worthwhile regardless of which architecture you choose.
What edge hardware actually constrains
Edge devices trade compute for power and cost, so the constraint is real and worth understanding before specifying.
| Dimension | Practical effect |
|---|---|
| Model size | Weights are usually quantised to INT8 to fit memory and accelerate inference. This costs a small amount of accuracy that must be measured, not assumed. |
| Throughput | Measured in frames per second across all streams, not per stream. A device handling 16 cameras at 5 fps each is doing 80 fps of work. |
| Frame rate | Most analytics do not need 25 fps. Sampling at 5 to 10 fps is common and cuts the compute load proportionally. |
| Model updates | Have to be pushed to every device rather than deployed once centrally. Fleet management becomes an operational requirement. |
| Failure domain | A device failure takes out one site rather than all sites. This is usually an advantage, but it needs a monitoring plan. |
Edge in a retrofit
The common misconception is that edge AI requires AI cameras. It does not. An appliance on site can pull RTSP streams from the existing NVR, which means the cameras and recorder already installed keep doing their job while the analytics layer is added alongside them. Where the estate supports ONVIF, discovery is faster.
This is the architecture that makes edge affordable on an existing estate, because the capital already spent on cameras is preserved. It is also why the capture layer and the intelligence layer are worth thinking about as separate decisions.
When cloud is the better answer
Edge is not universally correct, and vendors who claim it is are selling rather than advising. Cloud processing is the better choice when camera counts per site are low, when many small sites each need a handful of streams, when the output is analytical rather than real-time, or when your team already operates a cloud VMS competently. Adding an on-premise system that nobody on site can maintain is a real cost, and it is frequently larger than the bandwidth saving it produces.
The honest decision rule: choose edge when the decision is time-critical, the camera count per site is high enough that backhaul is expensive, or the footage cannot leave the premises. Otherwise cloud is fine, and hybrid, where detection runs locally and only flagged clips are uploaded for review, covers most of the remaining cases.
Frequently asked questions
Does edge AI require special AI cameras?
No. An on-site appliance can pull RTSP streams from cameras and NVRs already installed and run the analytics on those streams. AI cameras put the processing in the camera itself, which is a clean architecture for new builds, but it means replacing hardware. For an existing estate, an appliance alongside the current NVR preserves the capital already spent.
How much bandwidth does edge AI actually save?
For a 16 camera site at 2 Mbps per stream, sending video continuously means roughly 10.4 TB a month. Sending only detection results means well under a gigabyte. The ratio is around 17,000 to 1. The saving matters most where the uplink is a metered or leased connection, which is common at industrial sites.
Is edge AI less accurate than cloud AI?
Not inherently, but edge models are usually quantised to INT8 to fit the hardware, which costs a small amount of accuracy. Whether that cost is material depends on the task. Ask any vendor for the accuracy figure measured on the quantised model running on the actual device, not the figure for the full precision model on a server.
What happens if the edge device fails?
That site loses analytics until the device is restored, while other sites continue unaffected. Cameras and the NVR keep recording independently, so footage is not lost. This is a smaller failure domain than a central outage, but it does require device health monitoring so a silent failure is not discovered weeks later.
Does edge processing satisfy DPDP requirements by itself?
No. Processing on site reduces the number of parties handling personal data and makes retention easier to control, both of which help, but the obligations around notice, purpose limitation, retention and data principal rights apply regardless of architecture. Edge makes compliance easier to implement, not automatic. This is general information rather than legal advice.
How many cameras can one edge device handle?
It depends on the models running, the frame rate sampled and the resolution analysed, so a single number is meaningless. Ask for throughput in total frames per second across all streams for your specific model mix, then divide. A device rated for 16 streams running one model may handle six running four models.
Working out whether edge fits your site?
IndoAI runs analytics on site alongside your existing cameras and NVR. Send us your camera count, your uplink speed and what decisions the system has to make, and we will show you the arithmetic for your deployment.
Talk to an adviserReviewed by Dr. Vivek Gujar, Chief Strategy Officer at IndoAI Technologies Pvt. Ltd., a Pune-based edge AI camera platform founded in 2021. He reviews IndoAI's published technical claims for accuracy. Profile