Behavior detection on the cameras you own.
AIsle is a thin software layer between your existing CCTV system and your floor team. It captures behavior, not identity — and turns it into a phone notification before a theft happens.
- RT-DETR
- ByteTrack
- VideoMAE-v2
- Apache 2.0
- 30 fps
- < 2 sec
- RTSP / ONVIF
- None
From camera to staff phone, in under two seconds.
Ingest
RTSP / ONVIF feeds from existing IP cameras. ONVIF Profile S compliant. No NVR replacement.
Detect + Track
RT-DETR (Apache 2.0) for person detection at 30 fps; ByteTrack (MIT) for re-identification across the floor. No biometrics.
Behavior model
VideoMAE (Apache 2.0) action classifier flags concealment, sweeping, prolonged loitering, queue dwell, repeat-reach.
Alert
A 4-second clip + a one-line description lands on the floor manager's phone via SMS, Slack, or our app. One tap to acknowledge.
Eight behaviors, lots of false-positive engineering.
We tune for sensitivity per store. New customers default to conservative thresholds — better to miss one than to send a staff member after a tired shopper.
Concealment
Item moved into a bag, pocket, or under clothing — without payment behavior.
Sweeping
Multiple SKUs swept off a shelf in rapid succession.
Prolonged loitering
More than 30 seconds in one aisle without purchase intent (configurable).
Repeat-reach
Multiple visits to the same shelf, often signaling staging.
Cart abandonment
Cart left in an aisle by a person who exits without it — recoverable inventory.
Queue skip / dwell
Anomalous behavior near checkout — prep for a sweep or an exit-rush.
Bag-switching
Bringing in a bag, leaving with a different one, identified by shape/color only.
Coordination
Multiple people moving as a coordinated unit — a hallmark of organized retail crime.
One screen for the floor team.
Most days, your team won't open it — they'll act on phone alerts. The console is for shift managers, end-of-day review, and the rare deep-dive when an incident report is needed.
Store 04 — Kelowna West Grocery
6 cameras · 23 ms · today
- Today, 14:23CAM 03Concealment · Aisle 4 · 23 s loitering
- Today, 13:51CAM 05Sweep · repeated reach to high shelf
- Today, 13:11CAM 02Loitering · resolved by staff touch
- Today, 12:44CAM 03Cart abandonment · recovered
- Today, 11:03CAM 01Anomalous queue dwell
Edge-first by default. Cloud where it makes sense.
AIsle Edge
Default for Pro and Enterprise
- A small AIsle appliance (or a Linux box you provide) runs the model locally. Cloud sees only alert metadata + 4-second redacted clips.
- Required for cannabis and BC Liquor — keeps raw video on-store.
- Hardware: NVIDIA Jetson Orin or x86 with a single RTX-class GPU. We handle deployment.
- Setup: 2–3 hours including network walk-through.
AIsle Cloud
Starter and small-store Pro
- Camera streams encrypted in transit (TLS 1.3); inference runs in our Canadian region.
- Raw video retained for 24 hours, then auto-deleted. Alert clips retained 30 days.
- PIPEDA-compliant. Data residency: Canada-only.
- Setup: 1 hour for a 6-camera store.
What's under the hood.
| RT-DETR · Apache 2.0 | |
| ByteTrack · MIT | |
| VideoMAE-v2 fine-tuned on retail behavior · Apache 2.0 | |
| RTSP, ONVIF Profile S, RTMP | |
| 30 fps per camera (cloud), 15–30 fps (edge) | |
| < 2 seconds median frame-to-phone, < 5 seconds p99 | |
| Behavior-only · no facial recognition · no biometric extraction | |
| Cloud: 24h raw / 30d alert clips. Edge: configurable. | |
| PIPEDA · BC PIPA · SOC 2 Type I (Q4 2026), Type II (Q3 2027) | |
| Mobile app (iOS/Android) · SMS · Slack · Webhook · REST API | |
| Hikvision · Dahua · Axis · Avigilon · UniFi · NVR-agnostic |
Want a deeper technical walkthrough?
15 minutes with our founding engineer. We'll review your camera setup live.