CCTV/AI video analytics
AI video analytics · Melbourne & VictoriaYour cameras already see it. Now they can tell you.
We train detection models on your site and your problem — a hard hat missing at the gate, a person on the fence line at 2am, a car that isn't on the tenant list — and run them on the cameras you already have. Alarms land in the system you already use. Footage never leaves the building.
- Works with your existing CCTV — any RTSP or ONVIF camera
- Custom-trained detection, not a generic “motion” alert
- Processed on site — no cloud, no faces, no subscriptions per camera
Overlays shown as they appear in the operator view. Faces and plates are masked in these examples; on your site they are never analysed unless you ask for it.
Built-in camera analytics know “motion”. Yours will know your site.
Standard building blocks below, plus anything specific to your operation. If a person can spot it on the monitor, we can usually train a model to spot it for you.
People & vehicles
Reliable person, car, truck and bike detection that ignores wind, rain, headlights and animals.
Zones & line crossing
Draw a fence line or an exclusion zone on the phone. Alarm when it is crossed, by what, and in which direction.
PPE: hard hat & hi-vis
Check every person entering a zone for a hard hat and hi-vis. Log compliance; alert on a miss.
Plant & pedestrian proximity
Excavators, forklifts and trucks tracked as plant. Warn when a person gets inside the swing or travel path.
Counting & dwell
Entrance counts in and out, occupancy now, and how long people spend at a shelf, a queue or a door.
Number plates
Read plates at gates and car parks, match against an authorised list, open the boom or raise an alarm.
Loitering & after hours
A person in the yard at 2am, or lingering at the back door for five minutes, is a different event from a passer-by.
Thermal hotspots
On a thermal camera: a stockpile, battery store or pallet yard warming past a threshold, flagged hours before smoke. Early warning that supplements the fire alarm system.
Anything specific to you
A pallet left in a fire aisle, a gate left open, a trolley on the ramp, a delivery in the wrong bay. Tell us what it looks like and we train for it.
The same cameras, doing a different job on every site.
Below are the four kinds of site we are asked about most. Each one starts with the cameras that are already there.
Security at night, safety by day, one camera system.
Builders buy cameras for theft and after-hours intrusion. The same cameras can run PPE and exclusion-zone checks during the shift, and the security side stops false-alarming on tarps and wildlife.
- PPE at the gate and in live zones — hard hat and hi-vis checked on every person, compliance logged by the hour, misses sent to the supervisor.
- Exclusion zones around plant — a person inside an excavator's swing radius is an alert, not a near-miss report a week later.
- After-hours intrusion that is actually a person — verified before anyone is called out, so the monitoring bill stops climbing.
- Zones move with the site — redraw an exclusion zone from a phone when the crane pad moves. No retraining.
A fence line that only calls when it's a person.
Most perimeter systems are turned off within a year because they cry wolf. A trained detector on a visible or thermal camera tells a person climbing the fence from a fox, a plastic bag or a storm, and holds that judgement at 2am.
- Fence-line and yard zones armed on a schedule, with the direction of travel.
- Thermal and low light — the model is trained on your night footage, not a daytime demo reel.
- Verified alarms — a snapshot with the detection goes with every event, so the control room sees why it fired.
Count, dwell and loss prevention — without identifying anyone.
Foot traffic and dwell tell you whether the window display works and how many staff the floor needs. Behaviour-based loss prevention flags the situations that matter without facial recognition, which keeps you clear of the privacy findings that have caught larger retailers.
- Entrance counting and occupancy — in, out, and inside now, by hour and by door.
- Dwell and queue length at high-value shelves and the registers, with a staff alert when a queue builds.
- Loss prevention by behaviour — loitering, entry to staff-only areas, after-hours presence, and plates in the car park.
Forklift lanes, tenant lists, and what's happening out the front.
The same building blocks cover a lot of ground: a pedestrian in a forklift lane, a vehicle that isn't on the tenant list, or a simple count of people and cars passing a shop front over the week.
- Forklift and pedestrian separation — forklifts tracked as plant, lanes drawn on the floor, alert on a person inside.
- Car park plate matching — authorised list, visitor logging, unknown-vehicle and after-hours alerts.
- Vehicle attributes — body type and colour on every passing vehicle, with make where the camera angle allows.
Hotspot warning hours before there's smoke.
Resource-recovery sites in Victoria carry a fire risk management obligation under the EPA's combustible waste rules, and a pile can smoulder for hours before a smoke detector or a person notices. A thermal camera with a trained threshold model watches the stockpile temperature around the clock and alarms on a rise, not just a flame.
- Stockpiles, bunkers and battery stores monitored by zone, each with its own threshold and rate-of-rise alarm.
- False positives filtered — exhausts, sun-heated metal and plant working the pile are learned and ignored.
- A record for the regulator — temperature history per zone, exported for the fire risk management plan.
- Supplements, never replaces, the site's compliant fire detection and suppression. Alarms go to the site and its security system, not the fire brigade.
Why a model trained on your site beats the analytics in the camera.
Every modern camera ships with “smart” detection. It is trained on a generic dataset and tuned for nobody in particular, which is why it fires on the neighbour's cat and misses the person in dark clothing.
| Built-in camera analytics | Camera-trailer monitoring | PSC custom analytics | |
|---|---|---|---|
| Trained on your site's footage | No, generic model | No | Yes, and re-checked after install |
| Custom classes (PPE, plant, your own objects) | Person and vehicle only | No | Yes |
| False alarms from weather, animals, headlights | Frequent | Filtered by an operator, billed per event | Filtered by the model, before anyone is called |
| Works with the cameras you already own | Only that brand's cameras | Their trailer, their cameras | Any RTSP or ONVIF camera |
| Zones and rules editable by you, from a phone | Per camera, in the camera's own app | Call the provider | Yes, all cameras in one place |
| Footage stays on site | Often cloud-linked | Streamed to their control room | Yes, processed on an on-site box |
| Alarms into your security system or VMS | Sometimes | Separate system | Inception, Milestone, SMS, email |
| Ongoing cost | Usually none | Weekly hire plus per-event fees | Hardware once; optional support plan |
Built-in analytics are fine for a driveway. They stop being fine the first time an insurer or WorkSafe asks why an alarm didn't fire.
Three steps from your first footage to a live alarm.
Typical time from site visit to a working pilot is a few weeks, most of it waiting for enough footage of the thing you want to catch.
Capture footage on site
We connect to your existing cameras and record a few days of normal operation, day and night, from the angles that matter. Nothing leaves the site; we review it with you.
Site visit + 3 to 10 days of footageTrain and validate on your scenes
We label the events you care about, train a detection model for them, and test it against footage it has never seen. You get the hit rate and the false-alarm rate in writing before anything goes live.
Measured, not promisedDeploy on an on-site box, alarm into your system
A small processing unit sits on your camera network and runs the model on every stream. Alarms go to Inception, Milestone, a phone or an email, with a snapshot attached. Zones and schedules stay editable by you.
Live, with a 30-day tuning periodOn site, by design.
No cloud. No faces. No footage leaving the building.
The processing unit lives on your network. It looks at frames, decides whether an event happened, and keeps a snapshot only when one did.
- No facial recognition. Faces are never analysed, matched or stored. People are counted and tracked as “a person”, not as who they are.
- Plates only where you need them, at gates and car parks, held against your own list and nowhere else.
- Your footage is yours. Training clips are reviewed with you and deleted when the model is signed off.
- Signage and policy help so what you deploy lines up with the Privacy Act and your site's own rules.
Single-site unit
A fanless box on the camera network, sized to the stream count. Runs four to eight cameras per unit; stack units for more.
Solar and 4G for sites with no power
For construction and remote yards: cameras, processing and a 4G link on a solar trailer or pole, reporting back over an encrypted tunnel.
Thermal for fence lines and stockpiles
Where the perimeter is long or unlit, or a stockpile needs watching for heat, a thermal camera plus a model trained on thermal footage does the job at night without floodlights.
What site managers ask us first.
Will it work with the cameras I already have?
Do I need to replace my Milestone or Inception system?
How accurate is it, really?
Does it use facial recognition?
Is the thermal hotspot monitoring a fire detection system?
Where does the footage go?
What about a construction site with no power or internet?
How long does it take and what does it cost?
Which areas do you cover?
Tell us the one thing you wish your cameras would catch.
We'll look at your site and your existing cameras, and tell you plainly whether it can be done, what it would take, and what it would cost.
Detection rates quoted for any site are measured on that site's own footage during the pilot and stated in writing. Video analytics assist trained people; they do not replace a site's safety procedures, WorkSafe obligations or emergency services. Example images on this page are illustrative, with faces and plates masked. Peninsula Smart Care, 302 Charman Road, Cheltenham VIC 3192. See our privacy policy.