Platform Features — What the VisaRoxy Vision Layer Does

Q: How are false positives handled?

Three levers. The condition itself is written narrowly, for example a vehicle class rather than any moving object. A persistence threshold removes anything that appears for less than a set number of seconds. A deduplication window stops one real situation becoming forty events. Everything an agent raises is reviewable, so thresholds get tuned against real footage.

The product surface

Five stages, one pipeline per stream.

Each stage has a defined job and a defined output. Nothing is skipped and nothing is duplicated: a second agent on the same camera reuses the understanding stage rather than repeating it.

01 · Ingest: VisaRoxy pulls the stream from each camera and decodes frames at the rate the agent needs. Streams arrive over RTSP or ONVIF, so the camera model and vendor do not matter. A camera that already records to a VMS keeps recording; VisaRoxy reads a second stream from it.
02 · Understand: Each frame is resolved into the objects and areas the operation thinks in: people, vehicles, equipment, doors, and the zones drawn over the scene. Objects are tracked between frames, which is what allows the platform to talk about duration rather than presence — a pallet that has been in a lane for nine minutes, not a pallet that exists.
03 · Decide: The rule engine evaluates the agent's conditions against that structured output. A rule combines a condition, a persistence threshold, a schedule and a deduplication window. The output is either nothing, a log entry, or a named event with a severity.
04 · Act: Confirmed events are routed outward: an alert to a team, a request in a ticketing queue, an outbound webhook, a signal to connected equipment, or an escalation if nobody acknowledges within a configured window.
05 · Review: Every event is retained with its clip and its context, and can be searched by zone, event type and time range. Thresholds are tuned from this record, which is how an agent gets quieter and more accurate over its first weeks.

Capabilities

Grouped by what the platform has to get right.

Seeing, deciding and acting fail in different ways. These groups are how the product is built, and how a pilot is usually scoped.

Seeing

Scene understanding

The platform has to answer questions about a scene it has never been shown before, on cameras it has never seen, in lighting that changes through the day. Four things make that practical on a real site.

Object classesPeople, vehicles, equipment, doors, and custom classes trained for a site.
Zones and lanesPolygons drawn on the scene, with rules that apply only inside them.
Tracking over timeThe same object followed between frames, so duration is meaningful.
Camera calibrationPerspective handled so a zone drawn on screen matches the floor area it covers.

Deciding

Event reasoning

Understanding a scene is not the same as knowing whether something is wrong. The rule engine is where site knowledge is written down, and it is deliberately readable by the people who own the zone.

Conditions across objects and zonesA vehicle class inside a walkway, not any movement in a rectangle.
Persistence thresholdsAnything shorter than the threshold never becomes an event.
Schedules and operating hoursThe same rule behaves differently on a night shift or during a booked delivery window.
Deduplication and severityOne situation becomes one event, ranked so it reaches the right person.

Acting

Response and record

An event that nobody receives is a log line. Everything in this group exists to move a decision to the person or system that can do something about it, and to keep an attributable record of what happened.

Alerts with contextZone, event name, severity, duration and the clip around it.
Escalation pathsA second and third recipient if the first does not acknowledge.
Outbound automationWebhooks and a REST API into ticketing, dashboards or equipment.
Searchable historyEvents retained and queryable long after the clip window has closed.

Deep dive

The hard part is deciding without a human in the loop.

Detection is a solved category. The unsolved part is judgement: knowing which of the thousands of things happening in a frame deserves somebody's attention right now.

A rule in VisaRoxy is not a sensitivity slider. It is a sentence with four independent parts, and every part removes a category of false positive. Get the parts wrong and the agent becomes noise; get them right and it stays quiet until something is actually wrong.

Consider a walkway that runs between two dock doors. The site wants to know when a vehicle is inside it, because pedestrians use the same strip. Take the rule apart and each clause is doing specific work.

1
Condition — a vehicle class inside the walkway polygon

Not "movement in a rectangle". The condition names the object class and the zone, so a person walking through the walkway is not an event, and a shadow crossing it is not an event, because neither is a vehicle.
2
Threshold — present for longer than three seconds

A forklift crossing the strip at speed is a near-miss worth recording but not worth interrupting a shift for. Three seconds separates a passing wheel from a vehicle parked across the route.
3
Schedule — all hours, with a booked-delivery exception

The rule runs around the clock, but the yard team books a delivery window every Tuesday at 06:00 when the walkway is formally closed and vehicles are expected. The exception is written into the schedule rather than switched off by hand.
4
Deduplication — one event per vehicle per five minutes

A vehicle that stays in the walkway for twenty minutes should not produce four hundred events. The deduplication window keeps one event open and updates its duration until the vehicle leaves.

The result is an agent that raises a handful of events a week on a camera that used to produce hundreds of alerts a day. That difference is the whole product, and it comes from configuration rather than from a better model.

Vehicle in pedestrian walkway

Agent: Loading Dock Sentinel · Severity: high — safety

Live

Condition

object is vehicle inside walkway

Threshold

persists > 3 s

Schedule

all hours except Tue 06:00–07:00

Then

trigger alert notify Shift supervisor dedupe 5 min

Recent matches

Tue 08:41 Vehicle in pedestrian walkway · 11s Acknowledged in 2 min
Mon 17:03 Vehicle in pedestrian walkway · 4s Not acknowledged
Mon 09:22 Below threshold · 1.8s Discarded, not an event

Illustrative interface with sample data.

The workspace

Two screens do most of the work.

Configuration and review. There is no separate model-training tool to learn, because the platform's job is to apply your operational knowledge, not to ask you to build a model.

Agent builder

Dock 04 · Receiving

Cameras on this agent

Dock 01 — Entrance
Dock 02 — Shipping
Dock 04 — Receiving
Dock 03 — YardNot assigned; no rule applies here yet.

Operational rules

When dock congestion exceeds 8 minutes during operating hours, trigger an alert and notify the supervisor. When a vehicle enters the walkway, trigger a safety alert.

Escalation path

Shift supervisor
Site safety lead
Duty manager

Event review

Filter: safety severity · last 7 days · all zones

142 events

Awaiting acknowledgement

Today 08:41 Vehicle in pedestrian walkway Dock 04 · 11s · clip attached
Today 07:56 Pallet stacked above sightline Staging · 6 min · clip attached

Closed this week

Mon 17:03 Blocked emergency exit Cleared in 4 min · ticket 4412
Mon 14:19 PPE missing at Dock 02 Coaching logged · no ticket
Sun 22:04 Perimeter movement outside hours Verified as scheduled contractor

Illustrative interface with sample data.

End to end

One event, followed from frame to acknowledgement.

This is the same walkway rule, traced through the platform in the order the stages run. The whole sequence completes in seconds, not in a shift report.

One walkway event traced through the five platform stages.
Stage	What happens	What exists afterwards
Ingest	Camera 04 publishes a stream; the runtime decodes frames at the rate the agent needs.	Decoded frames. Nothing stored.
Understand	A vehicle is detected and tracked inside the walkway polygon across consecutive frames.	A tracked object with a class, a zone and a running duration.
Decide	Duration passes three seconds; the schedule allows it; the rule matches and opens an event.	One named event with severity, timestamp and zone. A dedupe window opens.
Act	The alert goes to the shift supervisor with the clip; the walkway beacon is signalled.	A delivered alert with an acknowledgement state and a delivery record.
Review	The supervisor acknowledges at 08:43. The event closes with its clip attached.	A searchable record: zone, event type, duration, response time, outcome.

What the record is worth

After a month, that history answers questions a walkthrough cannot: which zone raises the most events, which are acknowledged fastest, and which rules are producing noise that nobody acts on. Tuning the agent is a review of its own record, not a guess.

Technical questions

What engineers ask on the first call.

Architecture questions usually decide whether a pilot happens. These are the five that come up most often.

See the integration surface

Does every capability need its own camera or model?

No. One stream feeds every agent assigned to that camera. Scene understanding runs once per frame, and each agent reads the same structured output with its own watch list, thresholds and schedule. Adding a second agent to a zone does not add a second model or a second stream, which is why a site can grow its rule set without growing its infrastructure.

Can VisaRoxy run on the camera itself?

Some cameras ship an analytics slot that can host a limited model, and VisaRoxy can be scoped to use one where it exists. The full agent model, rule engine and event store run in the runtime you nominate, because the reasoning needs context across cameras and across time that a single camera does not have. A camera cannot know that a lane has been blocked for eight minutes if the vehicle entered from a different camera's view.

What happens when the network to the cloud drops?

An edge deployment keeps running. Inference, rule evaluation and event storage continue on the appliance, and events queue for delivery to outbound systems when connectivity returns. A cloud deployment depends on the link, which is one of the reasons edge is the common choice for sites with an unreliable WAN or a strict rule about video leaving the building.

How are false positives handled?

Three levers. The condition itself is written narrowly, using a vehicle class rather than any moving object. A persistence threshold removes anything that appears for less than a set number of seconds. A deduplication window stops one real situation becoming forty events. Everything an agent raises is reviewable in the event list, so thresholds get tuned against real footage during the pilot rather than guessed at the start.

Can we export events into our own data platform?

Yes. Events are available through the REST API and can be pushed to an outbound webhook as they are created, so they can land in a warehouse, a BI tool or a data lake alongside the rest of your operational data. Clip access is separate from event metadata and is controlled by the same access rules, so an analyst can query event counts without being granted access to video.

A vision layer, not another camera system.

Five stages, one pipeline per stream.

Grouped by what the platform has to get right.

Scene understanding

Event reasoning

Response and record

The hard part is deciding without a human in the loop.

Condition — a vehicle class inside the walkway polygon

Threshold — present for longer than three seconds

Schedule — all hours, with a booked-delivery exception

Deduplication — one event per vehicle per five minutes