beixiaocai/rebucca

What it solves

Rebucca is a multi-channel video access and intelligent control analysis platform. It solves the problem of integrating various video streams (like RTSP and GB28181) and applying automated AI analysis to detect specific events, such as intrusions or line-crossing, and triggering structured alarms based on those events.

How it works

The platform integrates video streams using ZLMediaKit and ONVIF discovery. It employs a two-stage analysis process: first, a "small model" (using YOLO via PyTorch, ONNX, or OpenVINO) performs initial detection; optionally, an OpenAI-compatible large model (LLM) is used for verification to increase accuracy. Users define "business algorithms" by drawing polygon regions on the video feed and setting rules (e.g., density, stay duration, or direction) to determine when an alarm should be triggered.

Who it’s for

It is designed for operators and developers needing a comprehensive surveillance system that combines traditional computer vision detection with modern LLM verification and customizable spatial alerting rules.

Highlights

• Multi-protocol Support: Compatible with GB28181 and RTSP streams.
• Hybrid AI Pipeline: Combines fast YOLO-based detection with optional LLM-based verification.
• Customizable Alerting: Supports five post-processing rules including intrusion, line-crossing, direction, density, and lingering.
• Operational Tools: Includes a control panel for monitoring, recording capabilities, and a statistical dashboard for alarm trends.
• Multilingual: Supports 7 different languages.