How AI Detects Shoplifting in Real Time

Traditional CCTV answers one question: what happened? AI-powered systems answer a far more important one: what's about to happen?

IntelliGuard uses multiple AI models working together in real time to detect concealment behaviour — the single most common theft method in Australian pharmacies. Here's how the pipeline works.

1. Person Detection & Tracking

The first layer uses state-of-the-art pose estimation to detect every person in the camera frame and track them across time. Each individual is assigned a unique ID that persists as they move through the store, even across partial occlusions. This is powered by YOLO-based detection running at over 30 frames per second on compact edge hardware — fast enough to never miss a frame.

2. Pose & Action Recognition

With each person tracked, the system extracts 18 skeletal keypoints per frame — wrists, elbows, shoulders, hips — and feeds sequences of these poses into a temporal action classifier. This model has been trained specifically on pharmacy concealment patterns:

• →Hand-to-pocket movement — the classic pocketing motion where a hand moves from shelf height to hip/pocket level
• →Hand-to-bag concealment — reaching into a personal bag while near a shelf, particularly with the body angled away from cameras
• →Clothing concealment — tucking items into jackets, waistbands, or under clothing layers
• →Product sweeping — rapidly gathering multiple items from a shelf into a bag

Unlike simple motion detection, these models understand the context of movement. Picking up a product to read the label looks very different from concealing it — even when the physical motion appears similar.

3. Object Detection (Hand-Bag-Pocket)

A specialised object detector identifies hands, bags, and pockets in the scene. When a hand enters a bag zone while near a shelf, the system correlates this with the pose data to confirm concealment. This multi-signal approach dramatically reduces false positives — a shopper putting their phone in their bag won't trigger an alert, but someone sliding a bottle of perfume in will.

4. Behavioural Scoring & Alert

All signals feed into a Finite State Machine (FSM) that maintains a threat score for each tracked person. The score rises with:

• →Sustained shelf interaction in high-value zones
• →Concealment-consistent hand trajectories
• →Dwell time exceeding normal shopping patterns
• →Repeated approach-retreat behaviour

When the score crosses the threshold, an alert fires. Staff receive an instant push notification with the suspect's face photo — captured by a separate face detection model — so they know exactly who to approach and where in the store they are.

5. All On-Premises, All in Real Time

The entire pipeline runs on a compact edge device in the pharmacy. No video is uploaded to the cloud. No third-party servers process your footage. Alert latency is approximately 2 seconds from detection to phone notification — fast enough for staff to intervene before the suspect reaches the exit.

This is not generic retail AI. IntelliGuard's models are purpose-built and continuously tuned for pharmacy concealment — the specific hand movements, shelf heights, product sizes, and store layouts that define cosmetics and fragrance theft in Australian pharmacies.

Explore IntelliGuard — AI theft prevention built for Australian pharmacies.