Can You Run AI Analytics on Your Existing CCTV? (Retrofit Guide)
Yes — if a camera or its DVR/NVR outputs a standard RTSP or ONVIF stream, AI analytics can read that feed and run detection on separate hardware or in the cloud, with no rip-and-replace. Retrofit fails only on too-low resolution, a bad sightline, or a recorder that can't expose a stream. The AI is a reader on your network; your existing recorder keeps recording in parallel.
That is the answer every vendor page gives you. It is also incomplete. Getting a stream out is necessary but not sufficient — the retrofit that actually breaks in production breaks on codec and compute, not protocol. This guide covers both: the obvious protocol layer, and the engineering layer nobody sells you.
Key takeaways
- Retrofit is a per-camera decision — most floors keep the large majority of cameras and swap only the few that fail on resolution or angle.
- Any camera or DVR exposing RTSP or ONVIF can feed AI; analog works via an RTSP-capable DVR or a per-channel encoder.
- The real trap isn't "can I get a stream" — it's "can my box decode that stream at scale." H.265 on an H.264-only box silently falls back to CPU and chokes.
- Record at 25–30 fps but infer at 5–10 fps — several-fold less compute, no loss of recorded footage.
- In India, buying new hardware after 1 April 2026 means STQC-certified, BIS-registered cameras — and for government/PSU buyers the software layer has its own STQC dimension.
Three questions that decide it
Before any vendor demo, answer these about your floor:
- Can I get a stream my box can actually decode — at scale? RTSP or ONVIF gets video off the camera. But if your cameras emit H.265 and your inference box only has hardware H.264 decode, every stream silently CPU-decodes and the box falls over before you've connected the whole floor. Match the codec to the decoder, not just the sticker.
- Is the picture good enough where detail matters? Pixel density at the target decides whether AI reads a helmet or a gauge — and there's a codified standard for it (below), not "higher res is better."
- Does the hardware stay legal to buy and the footage legal to use? In India, rules from 1 April 2026 govern what you can buy; the DPDP Act governs how you use footage today.
Clear all three and you retrofit. Fail one and you replace the specific cameras that fail — rarely the whole system.
How AI reads an existing camera — and where it chokes
AI video analytics doesn't need a "smart camera." It needs a video stream, delivered two standard ways:
- RTSP carries the live H.264/H.265 video — the media plane.
- ONVIF handles discovery, login and configuration — the control plane — then points the client at the RTSP stream. Most IP cameras of the last decade carry an ONVIF Profile S (adds H.264 live video) or Profile T (adds H.265 plus richer imaging and alarm/analytics events; a Profile-T device supports at least one of H.264 or H.265) conformance sticker.
The analytics box discovers the camera over ONVIF, opens the RTSP stream, decodes the frames, and runs its own models. The AI is a reader, not a replacement — your NVR keeps recording in parallel.
The decode trap. Decoding is the step the SERP skips. Modern GPUs decode video in dedicated silicon (NVDEC). If that block supports H.264 but the camera pushes H.265/HEVC, the software quietly falls back to CPU decode — and a handful of CPU-decoded 1080p streams saturate a box that would have carried dozens on the GPU. Cameras that pass the ONVIF-sticker test still fail here. In our own Mihnovka pilot we pull three 2 MP channels as H.265, 1080p at 25 fps over RTSP; the cameras even tag their HEVC as hev1, which QuickTime and Preview refuse to open until you re-tag to hvc1 (ffmpeg -c copy -tag:v hvc1). A codec gotcha, not a protocol one — exactly the class of problem a "just get RTSP" checklist misses.
One more catch: RTSP usually gives you video only. Event metadata the camera generates internally doesn't travel with the stream, which is why most retrofits run analytics from scratch rather than trusting built-in "smart" events. The ONVIF Profile M (published June 2021) standardises how analytics metadata is passed back, but support is thin on older fleets.
Mama (askthemama.com) reads exactly these RTSP/ONVIF feeds — point it at the floor and it returns a plain-language safety and efficiency summary, and a short phone walkthrough generates a camera placement plan before you buy any hardware.
What it costs to run — and how many cameras fit on one box
"No rip-and-replace" is a slogan until you cost the compute that replaces the cabling. Two numbers a plant head actually asks:
- Stream density. One modern datacentre-class inference GPU (e.g. NVIDIA L4) carries on the order of ~20–30 simultaneous 1080p analytics streams, depending on model and inference frame rate. A 60-camera plant is roughly a two- to three-GPU problem, not a per-camera one. Confirm streams-per-GPU for your model before assuming one edge box covers the floor.
- On-prem vs cloud. Amortised on an on-prem box, that GPU works out to single-digit dollars per camera per month; renting the equivalent GPU in the cloud runs several times higher, plus egress. For a fixed fleet that streams 24/7, on-prem edge almost always wins — the cloud's elasticity buys you nothing when load never varies. (Indicative — get live quotes.)
And because you reuse the cameras, go-live carries no cabling, mounting or construction — the slow, expensive part of a greenfield install.
The single biggest lever: infer slower than you record
Here's the trick that rescues a cheap or old-camera retrofit: decouple inference frame rate from recording frame rate. Keep recording at 25–30 fps for smooth playback and evidence, but run detection at 5–10 fps — plenty for people, forklifts, PPE and occupancy. That alone cuts analytics compute roughly 3–6× with zero loss of recorded footage, and it's what makes analysing a grainy DVR feed economically sane. No vendor page mentions it; it's the difference between "retrofit works on paper" and "retrofit pays."
Analog cameras and old DVRs aren't a dead end
Old analog (BNC) cameras on a DVR still have a path:
- DVR exposes RTSP per channel? Leave the cameras, put the DVR on the network, pull each channel into analytics.
- It doesn't? A video encoder digitises the analog signal into an ONVIF/RTSP stream — figure roughly INR 3,000–9,000 per channel, though quote it live; prices swing by channel count and brand. It extends the life of existing cabling and cameras.
Accuracy is still capped by the analog camera's real resolution: a grainy 960H feed can tell you a zone is occupied or a forklift is moving, but not read a gauge or recognise a face. Combine the encoder with 5–10 fps inference (above) and even a modest DVR fleet becomes affordable to analyse.
What resolution does AI video analytics need?
AI can only detect what the pixels show — how many pixels land on the target at a given distance. The codified reference is the DORI model (Detection, Observation, Recognition, Identification) in IEC 62676-4:2014, which sets a minimum pixel density in pixels per metre at the target for each task:
| Task | Min pixel density (IEC 62676-4:2014) | Informal US PPF |
|---|---|---|
| Detect a person is present | 25 px/m | ~8 PPF |
| Observe general activity | 62.5 px/m | ~15 PPF |
| Recognise a known individual | 125 px/m | ~20 PPF |
| Identify a person | 250 px/m | ~50+ PPF |
The standard is defined in px/m; PPF figures are informal US rules of thumb, not codified. The 2025 revision of IEC 62676-4 introduced a reworked framework, but DORI and the 25/125/250 px/m thresholds remain the widely-used industry reference.
A 2 MP camera on a wide aisle might give 8 PPF at the far end and 40 PPF up close — great for occupancy, useless for reading a label 15 m away. Two ways retrofits quietly die:
- Placement beats resolution. A well-aimed 2 MP camera outperforms a 4K one pointed at a pillar. If a structural column hides a work zone, no model recovers it.
- Sub-streams lie. Many DVRs default to a low-res sub-stream. Point analytics at the main stream where detail matters. (In our pilot, the main stream is channel
/101; the sub-stream/102looks fine on a phone and drops the detail you actually need.)
Does India's April 2026 STQC/BIS rule affect existing cameras?
If you're buying in India, one date changes your shopping list: 1 April 2026.
Under MeitY's Essential Requirements (Gazette notification of 9 April 2024) and the Compulsory Registration Order, CCTV cameras sold in India must be STQC-certified and BIS-registered for cybersecurity — secure firmware, encrypted comms, authentication. A MeitY Office Memorandum dated 16 January 2026 withdrew earlier sell-through exemptions, with no further extension past 1 April 2026.
| What | Effective | Hits already-installed cameras? | Action |
|---|---|---|---|
| STQC/BIS Compulsory Registration | 1 Apr 2026 | No — governs new sales only | Buy only certified models going forward |
| DPDP Act, 2023 (Rules notified 14 Nov 2025) | Phasing through mid-2027 | Yes — governs footage use now | Post signage, limit audio, avoid constant monitoring |
So your existing fleet is about retention — the mandate targets what can be sold. New or replacement cameras after the deadline should be on the certified list. Domestic brands (CP Plus, Prama, Matrix) certified early; Counterpoint Research puts domestic brands at ~80%+ of India's CCTV market as of February 2026. (Market estimate attributed to Counterpoint; verify against their report before quoting an exact figure.)
One nuance no retrofit guide flags: STQC certification isn't only a camera-hardware question. For government, PSU and critical-infrastructure procurement (GeM), the video software/VMS/analytics layer carries its own STQC dimension — so an AI layer added to your fleet is itself in scope for those buyers. Mama treats this as a first-class requirement for regulated Indian deployments rather than a hardware-only afterthought.
DPDP, in practice. CCTV footage is digital personal data under the DPDP Act, 2023. Safety and loss-prevention monitoring is generally permitted, but post clear signage, avoid constant intrusive monitoring, and be careful with audio — the habits are cheap to build now and expensive to retrofit.
FAQ
Can AI analytics work on any CCTV camera? Almost any camera that outputs RTSP or ONVIF and uses a codec your box can hardware-decode. Closed systems, images too low-res for the task, or an H.265 fleet feeding an H.264-only decoder are the exceptions.
Do I need to replace my DVR/NVR to add AI? Usually no. Analytics reads the live stream while the recorder keeps recording; you only replace it if it can't expose RTSP or ONVIF at all.
How many cameras can one analytics box handle? Order-of-magnitude, a modern inference GPU carries ~20–30 1080p streams — fewer for heavy models, more at 5–10 fps inference. Size streams-per-GPU for your model before assuming one box covers the plant.
Does India's April 2026 STQC rule affect my existing cameras? No — it governs cameras being sold, so your installed fleet is about retention. New purchases after 1 April 2026 should be STQC-certified and BIS-registered; for government/PSU buyers the analytics software has its own STQC dimension too.
