Erdős Vision Systems
Erdős Vision Systems
{{ bgCanvas }}

Cameras that finally understand.

Your CCTV, turned into a real-time intelligence layer. Never a black box.

Explore the system
See the architecture
HOW IT WORKS

From cameras to decisions.

One platform that sees, understands and acts. The moment it happens.

PERCEPTION
CAM 04LIVE

See

People, vehicles, hazards, events. On your existing cameras.

UNDERSTANDING

Analyse

Raw detections become flow, dwell, behaviour, anomalies.

ACTION
ALERT

Respond

Real-time alerts, with evidence. Act in the moment.

CAPABILITIES

One platform, many eyes.

One perception layer for safety, security and operations. Not just retail.

{{ it.n }} {{ it.label }}
Fire and smoke detection
01

Fire & smoke detection.

Catches ignition in the frame, before a ceiling sensor would.

No thermal hardware needed
Zone-level alerts on ignition
Tracks spread rate
Occupancy tracking
02

Occupancy tracking.

Live headcount and density, every floor and zone.

Live occupancy, per floor
Density and dwell, per zone
Footfall trends, full history
Vehicle and access monitoring
03

Vehicle & access monitoring.

Number-plate recognition and gate control, at every entry.

Number-plate recognition, every lane
Gate control by plate list
Full entry log: plate, time, authorisation
Intrusion detection
04

Intrusion detection.

Arms perimeters and restricted zones. Any breach, a real-time alert.

Per-zone perimeters, by schedule
Alerts on restricted or after-hours entry
Every alert carries a clip
Safety and fall detection
05

Safety & fall detection.

Flags missing PPE, unsafe behaviour, and falls the instant they happen.

PPE checks, live violation count
Fall detection from body pose
Unsafe-behaviour flags for review
Theft detection
06

Theft detection.

Surfaces concealment and shrink for a person to judge. Never accuses on its own.

Flags concealment and exit-dwell
A VLM confirms before alerting
Every flag goes to a human
THE SYSTEM

One engine underneath everything.

One shared perception engine. Every signal lands in a single operations view.

MODULE 01 · ENTRY / EXIT

Entrance-exit flow mapping

Every entry and exit, split workers vs. customers.

Live occupancy, per floor
Filterable history
After-hours alerts
Entry / exit detection
MODULE 02 · HEATMAP

Customer activity heat mapping

Floor-plan heatmaps: where customers gather, linger, and never go.

Interactive heatmap, time and day filters
Zone-level dwell rankings
Dead-zone and peak-hour views
Customer activity heat mapping
MODULE 03 · ACTIVITY

Worker activity level monitoring

See who is active, idle, or clustered, on every floor and shift.

Per-worker active, idle, clustered
Activity maps and shift summaries
Minimum-staffing alerts
Worker activity monitoring
MODULE 04 · ATTENDANCE

Worker attendance & escape detection

Highest sensitivity

Detects ghost clocking, tailgating, and unauthorised exits.

VLM-confirmed early-exit alerts
Daily integrity reports: present, ghost, absent
Tailgating logs with evidence
Attendance and escape detection
MODULE 05 · PERFORMANCE

Sales performance ranking

An objective engagement score that ranks your customer-facing staff.

Engagement leaderboards: daily, weekly, monthly
Per-worker interaction timelines
Zone-level engagement reports
Sales performance ranking
01 / 05
METHOD & DELIVERY

Validated on your footage before going live.

Each model is trained and tested on 30 days of your existing footage before any live inference runs.

01
WEEKS 1-3
Offline Validation
Ingest 30 days of your footage
Tune the models to your angles, lighting, and crowds
Calibrate every camera to your floor plan
Validate each module against ground truth
02
WEEKS 4-6
Live Inference
Switch to the live camera feed
Stand up the dashboard and alerts
Tune thresholds to live conditions
Hand off a system you can run
Launch day is tuning, not guesswork.
Modules validated4 of 4
Ground truth match98 percent
TECHNICAL DEPTH

Models tuned to your store.

A light model watches every frame; a heavier one reasons only when needed. Every layer is tuned to your floors, transparent and auditable. Never a sealed box you rent.

Perception

Sees and tracks every person, every floor

YOLO26RTMPOSEBOT-SORTCLIP-REID

Reasoning

Confirms what matters, in plain words

TRIGGERED VLMQWEN2.5-VL

Data & serving

Live state, queryable history, evidence

dataflow.txt
1detections -> kafka -> timescaledb
2evidence -> minio
3query -> fastapi
KAFKAREDISTIMESCALEDBMINIOFASTAPI

Runs anywhere

On-prem or cloud, no vendor lock-in

node
ON-PREM
CLOUD
ONNXTENSORRTDEEPSTREAMDOCKER
See the full technical stack +
PERCEPTION
YOLO26 detection
RTMPose keypoints
BoT-SORT tracking
CLIP-ReID re-identification
REASONING & SERVING
Triggered VLM · Qwen2.5-VL
Kafka · Redis · TimescaleDB
MinIO evidence store
FastAPI · ONNX · TensorRT · Docker
FAQ

The questions you're probably asking.

Hardware, privacy, deployment, ownership: the things that decide a pilot. Anything else, just ask.

Ask us anything
{{ f.n }} {{ f.q }} {{ f.sign }}

{{ f.a }}

Erdős Vision Systems
erdosvisionsystems.com
© Erdős Vision Systems 2026