yuturuishi/rebucca

What it solves

Rebucca is a real-time video surveillance and smart monitoring platform. It solves the problem of turning raw camera feeds into actionable, structured alarms. Instead of just recording video, it detects objects (people, vehicles, etc.), applies rule-based reasoning on regions you define (like intrusion or loitering), and generates reliable alerts with snapshots. It also addresses practical deployment pain: heavy detection models can be replaced with lightweight YOLO or D-FINE models, and an optional large language model can double-check detections to reduce false positives.

How it works

Rebucca pulls video streams using RTSP or GB28181 (with ZLMediaKit for forwarding) and optionally discovers cameras via ONVIF. The detection pipeline runs on small, fast models: a built-in D-FINE real-time end-to-end detector (exported to OpenVINO or ONNX Runtime) or YOLO in PyTorch/ONNX/OpenVINO formats. You can bind detection to drawn polygon zones and choose among seven post-processing rules: intrusion, line crossing, counting, direction, density, loitering, and absence. When a rule is triggered, the system produces an alarm with a snapshot image, and the alarm rate is throttled to match the detection frequency to avoid alert storms. An optional OpenAI-compatible large model can be used for a second-pass review to improve reliability. The whole flow is managed through a web control panel.

Who it's for

This is aimed at security and operations teams that need intelligent video monitoring: for example, guard rooms, cashier stations, construction sites, or server rooms. It's also suitable for integrators and developers deploying commercial surveillance systems, thanks to a permissive MIT license (with a note that the YOLO-PyTorch engine depends on AGPL-licensed Ultralytics unless you use the ONNX/OpenVINO path).

Highlights

• Supports multiple stream protocols: GB28181, RTSP, ONVIF discovery, with ZLMediaKit relaying.
• Includes D-FINE as a built-in detection framework (ICLR 2025), offering high localization accuracy and low latency without NMS.
• Dual inference engines: OpenVINO 2024.6 and ONNX Runtime 1.30.0, both CPU and GPU capable.
• Seven rule-based post-processing behaviors for flexible zone alarm scenarios.
• Every alarm includes a snapshot; no more 'empty' alarms without evidence.
• Alarm frequency strictly throttled to the detection frequency, preventing alert storms.
• Chinese label rendering for alarm snapshots, avoiding garbled characters.
• Optional large-model review for reducing false positives.
• Web control panel with dashboards, statistics, and a one-click restart.
• 7-language UI support and MIT-licensed source code.