How to Use CCTV for Retail Analytics: From Security Footage to Store Intelligence

Start with what the cameras were installed for

Security cameras are mounted to see faces at the door and hands at the till. Analytics wants the opposite: a high, steep view of the floor where people do not hide behind each other. A typical Indian showroom or EBO has two to eight cameras on a DVR or NVR, mounted at 2.4 to 3 metres, one or two at counter height, often with a wide lens that bends the edges of the frame.

That setup gives you an entrance count that is usually trustworthy, zone dwell in open areas that is fair, and path tracking across the whole floor that is poor, because tracks break every time someone passes behind a pillar, a display or another shopper. Knowing this before you start saves the most common mistake, which is promising the network a heatmap the cameras cannot produce.

The method in six steps

1. Inventory every camera: view, resolution, mount height, whether the recorder exposes a stream (RTSP or ONVIF), and whether the lens is wide or normal.
2. Choose measures the views can defend. Entrance count and counter wait come first. Zone dwell where a camera looks down on the zone. Skip full paths unless a camera is top-down.
3. Put a small edge computer in the store that reads the streams. Frames are processed there and never leave the store; only events do. This is the DPDP-safe design and also the cheap one, because no video crosses the network.
4. Draw lines and zones on each camera view: an entrance line, display zones, the counter, the waiting area, the staff-only area used to exclude staff.
5. Validate for a week by hand. Count the door for two hours on two days and compare. Time five queues with a phone. If the numbers differ by more than 10 to 15 per cent, fix the camera or the line before trusting anything downstream.
6. Read the events by hour beside POS and the sales conversation. Footfall alone is a denominator. It becomes analytics only when something explains it.

What an existing CCTV setup can and cannot measure

Where accuracy goes wrong

• Occlusion: at low mounts one shopper hides another and the count drops.
• Crossing: two people cross and the tracker swaps their IDs, so one long dwell becomes two short ones.
• Glass and evening light: reflections in a showroom window create ghost detections; a dim floor at 8 pm loses real ones.
• Staff counted as customers, unless they are excluded by a staff zone, a uniform colour rule or a clock-in pattern.
• Double counting at the door when a shopper steps out to take a call and comes back.
• Wide-angle lenses stretch people at the edges of the frame, where detectors are weakest.

Reading the events against the conversation

The camera records what happened around a sale: how many came, how long they waited, whether anyone reached them, where they lingered. The consented conversation records what happened inside it: the objection, the offer, the ask for a number. Neither explains conversion alone. Joined by time and zone, an unattended walk-out from the premium zone at 6:40 pm read next to the conversations running at 6:40 pm, they do. Joined by identity they would be unlawful and, for the purpose, unnecessary.

Borentis today captures consented in-store conversations on the advisor's phone, in Hindi, English and Hinglish, and scores them on the playbook. ShopperDNA, its vision product, is on the roadmap: it would join camera events such as footfall, group size, wait time, unattended walk-outs and display hotspots to those conversations by time and zone.

Privacy before the first frame

• No facial recognition, no age or gender estimation, no emotion reading. Behavioural events, not people.
• A notice at the entrance that cameras are used for security and for anonymous counting, in the local language and English.
• Footage retention stays whatever the security policy says; the analytics layer keeps counts and timestamps, not clips.
• Edge processing so no frame leaves the premises, which keeps the data-transfer question simple under India's DPDP Act.