Suspicious Activity Detection

GitHub project

The Suspicious Activity Detection (SAD) application is a store-wide loss prevention reference workload that demonstrates how Intel® SceneScape, OpenVINO™, and a Vision Language Model (VLM) can be combined on a single Intel® platform to detect suspicious in-store behavior in real time.

It combines several services:

• swlp-service: MQTT-driven core that subscribes to SceneScape, manages per-person session state, evaluates declarative rules, and emits alerts.

• Behavioral Analysis Service: Runs pose detection plus a VLM (Qwen2.5-VL) to confirm whether a person is concealing merchandise in a HIGH_VALUE zone.

• Alert Service: Time-window-based deduplication and downstream delivery of alerts to MQTT topics, REST consumers, and the dashboard.

• Frame storage (SeaweedFS / MinIO): Cropped person frames stored per visit; a per-alert evidence prefix is created when an alert fires.

• Gradio UI: Dashboard for live alerts, sessions, and evidence frames.

Together, these components illustrate how vision-based AI inference, rule evaluation, and downstream alerting can be orchestrated, monitored, and visualized for a retail loss-prevention scenario.

Supporting Resources

• System Requirements – Hardware, software, and network requirements, plus an overview of the AI models used by each workload.

• Get Started – Step-by-step instructions to build and run the application using make and Docker.

• How It Works – High-level architecture, service responsibilities, and data/control flows.

• How-to Guides – Guides on key topics like SceneScape setup, pattern authoring, and benchmarking.

• Release Notes – Version history and known issues.

Disclaimer: This application is provided for development and evaluation purposes only and is not intended for production loss-prevention or surveillance use without further validation, deployment hardening, and compliance review.