Integration Guide · Edge AI · India

How to Integrate AI Analytics with Any CCTV System in India

You integrate AI analytics with existing CCTV in India by pulling each camera's RTSP stream — discovered via ONVIF — into an on-site edge AI appliance, with no camera replacement and no cloud dependency. Choose direct-to-camera, NVR-relay, or switch-mirror access per camera, keep processing on-prem, and tune streams to protect bandwidth.

IndoAI Field Engineering · 25 June 2026 · 14 min read

Why AI analytics CCTV integration matters in India

Most real-world CCTV deployments in India are assembled systems, not single-vendor stacks. One camera brand sits at the gate, another inside the building, a third handles PTZ, and the NVR arrived from whoever won the tender. That is normal — but it means your upgrade path breaks the moment you assume a clean, uniform estate.

Two open standards make interoperability realistic across that mess:

For an India-first buyer, on-prem analytics is not a preference — it is usually a constraint. Multi-site bandwidth is limited and links are unstable; privacy and compliance obligations apply to identifiable video; and security teams need alerts even when the internet is down. Whether you are running retail loss-prevention or factory safety analytics, the box has to keep working when the WAN does not.

RTSP and ONVIF: the two standards that make it possible

It helps to keep these two separate in your head. RTSP is about streaming — control and transport of the video itself. ONVIF is broader: it handles device discovery, configuration management, time sync, events, and standardised access to streams across vendors. Used well, ONVIF turns "go find the URL for each camera by hand" into a structured onboarding step.

RTSP vs ONVIF in AI analytics CCTV integration: RTSP transports the video stream over port 554 while ONVIF discovers cameras and NVRs and fetches stream URIs, both feeding an on-prem AI edge box.
Figure 1 — RTSP carries the stream; ONVIF discovers devices and stream options. They complement, not replace, each other.
RTSP vs ONVIF — what each one actually does
DimensionRTSPONVIF
Primary jobLive stream transport/controlDiscovery, config, streaming access
Default port554 (sometimes 8554)HTTP/HTTPS per configuration
Best forThe "works-everywhere" baselineScaling onboarding across mixed brands
Watch-outURLs vary by vendor; may be undocumentedOften shipped disabled; time drift breaks events
Industry update ONVIF has communicated deprecation guidance around Profile S and recommends moving toward Profile T with stronger security such as TLS/HTTPS. Deployed Profile S systems keep operating, so in the Indian field you will see both S and T. A good integration supports both.

Three integration patterns that work — pick one

There are three practical ways to feed an edge AI appliance from existing CCTV. The "best" one depends on your NVR/VMS and how much control you have over the cameras.

Pattern A · Recommended

Camera streams directly to the AI box

Path: Camera (RTSP/ONVIF) → AI box → events/alerts + optional clips → security team.

Pros: cleanest, lowest latency, easiest to scale analytics.

Cons: you must hold camera credentials and ensure network reachability.

Pattern B · Cameras locked

NVR provides streams to the AI box

Path: Camera → NVR → AI box pulls streams from NVR (RTSP/SDK/secondary stream).

Pros: avoids touching camera configs.

Cons: some NVRs expose limited streams; you may get lower quality or higher latency.

Pattern C · Special cases only

SPAN / mirror from the switch

Path: Camera traffic mirrored at the switch → AI box.

Pros: useful when credentials are unavailable.

Cons: messy and brittle — avoid unless you have strong networking support.

Three AI analytics CCTV integration patterns side by side: Pattern A streams directly from cameras to the AI edge box, Pattern B routes through the NVR, Pattern C buffers via NVR for review.
Figure 2 — Patterns A, B and C: direct-to-camera, NVR-relay and buffered review, each feeding the same on-site edge box.

Most Indian deployments settle on a hybrid: Pattern A for the cameras that earn their keep (gate, billing counters, perimeter) and Pattern B for the rest.

A field-tested, step-by-step integration guide

The following six-step sequence is how clean rollouts actually happen. Skip the early steps and you pay for it in week three.

Step 1 — Build a camera inventory that's actually usable

Create a sheet capturing, for every camera: make/model, location, purpose (gate, cash counter, warehouse aisle); IP address, VLAN/subnet, PoE switch port; resolution, FPS, codec (H.264/H.265), main and sub-stream; and access method (RTSP URL known? ONVIF enabled? NVR-only?).

This is the first place projects fail. If you cannot answer "which stream do we pull, and from where," integration degrades into trial-and-error.

Step 2 — Decide which stream you will use (main vs sub)

Analytics does not always need your heaviest feed. Many sites overload the network by pulling the main stream for everything when the sub-stream would do.

Main stream vs sub-stream selection guide for AI analytics: main stream for face recognition and ANPR, sub-stream for intrusion, people counting and queue monitoring, with a quick task-to-stream lookup table.
Figure 3 — Main stream buys detail for identity tasks; sub-stream buys scale for broad detection. Start sub, promote where detail matters.
Practical stream choice by analytics task
Analytics taskRecommended streamWhy
People count, queue, shoplifting, intrusionSub-stream often sufficientLower bitrate; fine if placement is good
Face recognition, number plate (ANPR)Main stream / high-quality subNeeds more pixels-on-target for accuracy

To size bandwidth realistically, use vendor bitrate guidance. Typical tables show 1080p around 2048 Kbps for H.265 at 30 fps, while 1080p H.264 max bitrate can reach 4096 Kbps at 30 fps — values shift with settings and scene complexity. The translation is blunt: 16 cameras can be "fine" or "painful" depending entirely on stream choice.

Step 3 — Plan network and security (in India, most issues are networking)

Minimum best practice for any serious SME or enterprise:

Network segmentation for AI analytics CCTV integration: cameras and NVR on a dedicated CCTV VLAN, the AI edge box on its own VLAN behind a firewall allowing only RTSP and ONVIF, and the corporate LAN kept separate, with cloud optional.
Figure 4 — A safe topology: CCTV VLAN, isolated AI-box VLAN behind a firewall (RTSP + ONVIF allowlist), corporate LAN separated, cloud optional.

If you deploy face recognition — or anything that can identify a person — treat it as personal data processing. India's DPDP Act creates rights around access, correction and erasure, and frames obligations on processing. With rules now notified, operational compliance is becoming concrete rather than theoretical.

Step 4 — RTSP integration (the works-everywhere baseline)

Most cameras and most NVRs can hand you RTSP. You need the RTSP URL format (vendor-specific), a username/password, and a decision on TCP vs UDP transport — TCP is more firewall-friendly. RTSP's default port is typically 554; alternate ports like 8554 appear when 554 is blocked or repurposed.

Practical checklist: confirm you can open the stream from the AI box network (not your laptop Wi-Fi); prefer the sub-stream for bulk analytics; and lock codec and FPS to stable values rather than "auto," which causes bitrate spikes.

Common India failure modes NVR and cameras behind double NAT or mismatched subnets across multi-building campuses; a vendor-specific "stream encryption" toggle that breaks third-party RTSP clients; and the classic — a channel partner changed passwords and never documented them.

Step 5 — ONVIF integration (the scale-it-properly layer)

ONVIF is more than a checkbox. Used well it lets you discover devices, pull supported profiles, confirm stream URIs, and manage time sync and events consistently. Profile S devices send video over IP to Profile S clients that configure and request streaming; Profile T extends this with stronger security.

ONVIF checklist: enable ONVIF on the camera/NVR (many ship with it off); ensure system time is correct (time drift breaks event correlation); and prefer secure auth over weak legacy modes.

Step 6 — NVR/DVR realities (what you can and cannot assume)

In Indian deployments, "DVR" usually means analog cameras via an encoder, while "NVR" means IP. That changes your options.

Rule of thumb: for high-value outcomes (face recognition, ANPR, shoplifting) direct camera access is worth the effort; for broad monitoring (intrusion, loitering, crowd) NVR channel streams are usually acceptable.

When the stream won't play: a field troubleshooting path

Most "it doesn't work" calls in India are networking, not the analytics. Work this decision path top-down before you blame the AI box — ping reachability first, then credentials, then port, then URL, then ONVIF, then codec, then NAT.

Troubleshooting decision tree for an AI analytics CCTV integration stream that will not play: check ping reachability, credentials, RTSP port 554, RTSP URL, ONVIF discovery, codec compatibility and NAT, with common India field root causes.
Figure 5 — Stream-not-playing decision tree. Most failures resolve at reachability, credentials or port before you reach codec or NAT.

A worked bandwidth example (so 16 cameras don't surprise you)

Worked example · 16-camera retail store

Put 12 cameras on sub-stream at ~0.6 Mbps each for general analytics, and 4 cameras on main stream at ~3.5 Mbps each for the billing counters and entrance.

(12 × 0.6) + (4 × 3.5) = 7.2 + 14 = ~21.2 Mbps of analytics ingest on a dedicated VLAN.

Flip those 12 cameras to main stream and the same site jumps toward ~56 Mbps — the difference between a stable rollout and a network nobody trusts.

Three example deployments you can sanity-check

Recommended stream plan by camera count for on-prem AI analytics: small retail 16 cameras (4 main / 12 sub), housing society 64 cameras (8–12 main), factory 120 cameras (15–25 main), with use cases per stream.
Figure 6 — Stream mix by deployment size. Numbers scale, but the principle holds: sub-stream for breadth, main stream for identity and fine detail.

Retail store · 16 cameras · shoplifting + queue + people count

Use the sub-stream for the 12 general cameras and the main stream for the 4 high-precision ones (billing, entrance). Per-camera bitrate varies by codec and FPS — roughly 2 Mbps (H.265) to 4 Mbps (H.264 max) at 1080p. Outcome: stable analytics with no WAN upgrade, alerts pushed to WhatsApp or app.

Housing society · 64 cameras · gate face recognition + ANPR + intrusion

Gate cameras run on the main stream with tuned placement, a dedicated VLAN and strict access control. Perimeter cameras run sub-stream analytics with schedule-based intrusion. Keep all processing on-prem and store only the events and clips you actually need.

Factory · 120 cameras · PPE + forklift near-miss zones

Segment cameras by use-case: safety-zone cameras are prioritised, the rest run on a lower stream. Use ONVIF discovery to speed onboarding across mixed brands, then run a "proof-of-coverage" test measuring false positives by zone and lighting shift before you scale.

What "good" looks like after integration

A professional rollout ends with four artefacts: a documented camera-to-model mapping (which analytics runs where); a bandwidth and compute budget that stays stable; an operations rhythm of alerts, review and continuous tuning; and a change-control process so adding cameras or models does not break the system.

This is exactly where the platform approach pays off. Once streams are integrated, adding a new analytic should feel like installing a capability, not starting another integration project. That is the design principle behind IndoAI's programmable AI camera platform — and it is why a second or third use-case takes days, not another procurement cycle.

Architecture at a glance

Compliance: identifiable video and the DPDP Act

The moment a deployment can identify a person — face recognition being the obvious case — the footage becomes personal data, and India's DPDP framework applies. Build the controls in from day one: defined retention windows, restricted access to identifiable feeds, and a defensible answer for access, correction and erasure requests. An on-prem, edge-first design makes this materially easier, because the identifiable data stays inside your network and under your retention policy rather than scattering across third-party clouds.

FAQ

Frequently asked questions

Can I add AI analytics without replacing my cameras?

Yes. If your cameras or NVR can expose RTSP streams — and ideally ONVIF for discovery and control — you can route them into an edge AI appliance without replacing hardware. RTSP is widely supported and uses port 554 by default.

What if I don't know my RTSP URLs?

ONVIF discovery can often surface stream URIs and supported capabilities, reducing vendor-specific guesswork when RTSP URLs are undocumented.

My setup is mixed brand. Will it still work?

Mixed brand is exactly why RTSP and ONVIF exist. ONVIF cites interoperability across more than 33,000 conformant devices and clients over time — that is what makes a multi-vendor estate integrable.

Do I need internet for AI analytics?

Not necessarily. On-prem analytics can run fully locally. Internet is only needed for remote access, software updates, or optional cloud reporting.

What ports do I need to open?

RTSP commonly uses port 554 by default, with alternates like 8554 in some setups. ONVIF uses HTTP or HTTPS depending on configuration. Open only what the appliance needs to reach the camera subnet.

How much bandwidth will this consume?

It depends on resolution, codec, FPS and scene motion. Vendor tables show 1080p ranging from about 2 Mbps (H.265) to 4 Mbps (H.264 max). Use the sub-stream for wide-coverage analytics and reserve the main stream for high-precision tasks.

Will my existing NVR be impacted?

If you pull camera streams directly, the NVR is unaffected. If you pull from the NVR, confirm it supports additional clients without choking, and run a 48-hour soak test.

DVR with analog cameras: can I still do analytics?

Yes, but you typically analyse per-channel encoded streams from the DVR or NVR rather than individual camera ONVIF feeds. Quality depends on encoder settings.

What is the difference between RTSP and ONVIF?

RTSP is primarily for streaming control and transport. ONVIF is broader: device discovery, configuration management, and standardised streaming access across vendors.

Is ONVIF Profile S enough?

Profile S is widely deployed and still operates in the field. ONVIF now recommends moving toward Profile T and stronger security such as TLS/HTTPS, so a good integration should support both S and T.

How do I keep this secure?

Segment cameras onto a dedicated VLAN, restrict access, use strong passwords, disable unused services, and prefer secure authentication. For identifiable video, design retention and access controls carefully.

Face recognition on CCTV: any compliance concerns in India?

Yes. If the video is identifiable personal data, India's DPDP Act makes rights and obligations relevant — including access, correction and erasure. Treat face recognition as personal data processing from day one.

How do I avoid false alarms for intrusion or loitering?

Use camera-specific zones, schedule policies, and thresholds tuned to time-of-day lighting. Start with a pilot on 5–10 cameras and expand once false positives are acceptable.

How long does a typical integration take?

For 16–32 cameras with known credentials and clean networking, a few days including testing. For 100+ cameras with mixed brands and undocumented networks, plan a structured discovery and phased rollout.

What should I insist on from my installer or system integrator?

A complete camera inventory, documented credentials handover, a VLAN plan, a stream plan (main vs sub), and a post-deployment tuning schedule.

Plan your integration

Not sure which pattern fits your site?

Get a tailored blueprint for integrating AI analytics into your existing CCTV — stream plan, network design and use-case mapping, built around the cameras you already own.

Get your free site assessment
I
IndoAI Field Engineering
Edge AI deployment & systems integration

IndoAI's field engineering team designs and commissions edge AI deployments across retail, industrial and residential sites in India — integrating analytics into existing CCTV without rip-and-replace, edge-first and DPDP-aware.