A full-stack security system spanning visual detection, business rules, persistence, live delivery and human response. It accepts on-demand images or continuously watches image and video folders, turns YOLOv5 detections into actionable alerts, and synchronizes every state change across Web and HarmonyOS clients.
This is a full-stack security platform that turns object detection into actionable alerts, closing the loop from AI recognition through rule mapping and database persistence to WebSocket push and front-end handling. Recognition has two entry points: entering an image URL by hand, or dropping images and videos into the ai-watch folder, where a background thread scans every 2 seconds to stand in for a camera. Spring Boot, MySQL and WebSocket handle state and real-time sync, while the web and HarmonyOS clients are used for review and handling.
If detection results in a surveillance setting stop at bounding boxes, they cannot support later handling. This project maps model output to business alerts and attaches on-scene evidence and a handling state to each one, forming a traceable record.
The manual entry forwards an image URL to the AI service; the directory entry has a background thread scan ai-watch every 2 seconds, push results back and archive source files into processed. Detection results are mapped by rules to smoke/fire, crowd gathering, intrusion, vehicle, animal or object anomalies, and only broadcast over WebSocket once the transaction has committed.
Inference runs on FastAPI with YOLOv5 (both a COCO model and a fire/smoke model); the business side is Spring Boot 3.3 with MySQL 8, using WebSocket for real-time push. Interfaces are built with Vue 3 and ArkTS / ArkWeb. Vue is served as a local static build directly by Spring Boot, so the project runs straight after cloning; a one-command Docker deployment is also provided.
Five pages are implemented — dashboard, devices, AI recognition records, live alert centre and handling history — together with a code walkthrough tutorial. Actual detection quality should be assessed against the dataset, confidence threshold and deployment environment in use.
The admin front end is served as static files directly by Spring Boot, so the project runs straight after cloning with no separate Node.js environment.
Every step from “what the model saw” to a completed response preserves state and visual evidence.
Network images, uploads and sampled video frames are supported, with multi-frame results merged into one event.
COCO and fire/smoke models cover people, vehicles, objects, animals, smoke and flames.
Detected classes become domain alerts such as intrusion, crowding, fire, vehicle or object anomalies.
WebSocket updates are emitted after the database transaction commits, keeping screens consistent.
Alerts move from pending to in progress to closed, with invalid repeat actions rejected.
Annotated detections and video contact sheets remain attached to their alert records.
Manual submissions and the ai-watch folder share the same backend workflow.
Image URL, uploaded file, or watched image/video.
Inference, annotation and frame aggregation.
Mapping, validation and transaction control.
Devices, alerts, states and evidence.
Live events reach every connected console.
Operators can review the boxes behind an alert instead of trusting an opaque result.
The mobile client combines ArkWeb with device-aware interaction and native sharing.
Connect by LAN address with navigation, loading progress and demo-session sign-in.
The floating action control adapts to left- or right-hand holding status.
Tap-to-share and air gestures transfer a controlled demo session to nearby devices.
A coordinated browser, business, vision, data and device stack.