// INSIGHTS
CCTV and AI-Assisted Monitoring Across a Multi-Venue Event
6 min read
A five-day business summit runs across six venues in one city. Three hotels, a convention centre, a heritage property and an offsite dinner venue. Between them, 340 cameras from four different manufacturers, three of which were installed by the property and one of which is ours. The client wants a single command centre. The real question is not whether the feeds can be aggregated. It is what a human being in that room can usefully do with them.
A wall of video is not monitoring
Research on control room performance has been consistent for decades. Operator detection rates for events on unattended monitors fall sharply after about 20 minutes, and degrade further as the number of simultaneous feeds rises past roughly nine to sixteen. A 340-camera wall does not give you 340 cameras of coverage. It gives you a room full of screens that nobody is meaningfully watching and a recording archive you review after something has already happened.
The useful design inverts this. The default view is not video. It is a list of events, ordered by priority, each of which opens the relevant camera at the relevant timestamp. Video analytics run continuously at the edge and raise events. Operators work the event queue. The wall exists for situational context and for the two or three feeds that genuinely warrant persistent attention, such as a main entrance during peak arrival.
What raises reliable events at an event of this type is a short list. Line crossing into a restricted area. Loitering in a service corridor. Unattended object in a static zone. Camera tamper or signal loss, which is the most underrated alert in the category. Queue length at a defined portal. Vehicle presence in a fire lane. These are geometric and temporal rules over detections. They work because they are narrow.
What the estate looks like in practice
How the platform is built
Vendor-neutral ingest
Standard ONVIF and RTSP streams are pulled from existing NVRs and cameras. We do not require a rip-and-replace of property-owned infrastructure to get analytics running.
Inference at the venue edge
Detection runs on a local edge node at each venue. Only events, thumbnails and requested clips cross the link, which keeps a six-venue deployment viable on ordinary business broadband.
Event-first operator interface
A prioritised, filterable event queue with one-click jump to the source clip. Every event carries a state, an owner and a resolution note.
Audit log and retention control
Every view, export and search is logged against a named operator. Retention periods are set per venue and enforced automatically rather than left to whoever remembers.
From estate audit to live operation
Camera and network audit
We enumerate every camera, its resolution, codec, mount height and field of view, and mark which are analytics-capable. Typically 40 to 60 percent of an inherited estate is usable for analytics as mounted.
Bandwidth and edge sizing
Uplink capacity at each venue determines how much runs locally. We size edge nodes to the camera count and the analytics selected, with headroom for the busiest day.
A tuning week before the event
Zones, thresholds and schedules are tuned against real footage from the actual venue. Untuned analytics on day one produce hundreds of events per hour and destroy operator trust permanently.
Shift design and written SOPs
Each event type has a documented response, an owner and an escalation path. Shifts are structured around attention limits with rotation and breaks built in.
What usually goes wrong
False positives kill these deployments. A tree moving in wind, a reflective floor, a banner flapping over a tripwire, a housekeeping trolley parked in an unattended-object zone. Untuned, a 340-camera estate can generate several hundred events an hour. Operators learn within one shift to dismiss everything without looking, and from that point the system is worse than no system, because it manufactures a false sense of coverage. Tuning is not a nice-to-have. It is the majority of the deployment work. Model drift is the slower version of the same problem: lighting rigs change, layouts change between days, and thresholds that were correct on Monday are wrong by Wednesday.
We also decline some requests. Behaviour-prediction analytics that claim to flag suspicious intent, aggression or emotional state have weak published validation and well-documented demographic bias. Deploying them at an Indian venue means acting on a machine's guess about a person, disproportionately against some groups, with no meaningful recourse for the person flagged. We do not sell them. On face recognition across general CCTV, our default is no. It converts a security estate into an identity surveillance system, and under the DPDP Act it changes the purpose, the notice obligation and the risk profile entirely. Where a client has a genuine and lawful need, it should be confined to specific cameras at specific access points, with signage, a defined retention period and a named data fiduciary, not applied across the estate because the capability exists.
Common questions
Can you use the venue's existing cameras?
Usually a substantial share of them. Anything exposing a standard ONVIF or RTSP stream can be ingested. The limit is physical: cameras mounted low and oblique for face capture perform poorly for zone and counting analytics. The audit tells you exactly which of your cameras fall into which category.
How much bandwidth does a multi-venue setup need?
Far less than streaming every feed centrally. With edge inference at each venue, a site with 60 cameras typically needs 10 to 20 Mbps of stable uplink for events, thumbnails and on-demand clip retrieval. Full centralised streaming of 60 feeds would need roughly ten times that.
Do you run facial recognition on public CCTV feeds?
Not by default, and we will argue against it. Estate-wide face recognition is disproportionate for almost every event security requirement and creates significant DPDP exposure. Where identity verification is genuinely needed, we confine it to controlled access points with clear notice and consent.
Audit the estate before you buy the analytics
Send us your camera list and venue map and we will tell you what is usable today and what needs changing.
