Hardware
Beyond software
Float:Drone
Real-time object detection, tracking & avoidance drone system
FIG.01 — ASSEMBLY
FIG.02 — FLIGHT CONTROL
FLOAT:DRONE is an intelligent flight platform powered by Jetson, using camera feeds and sensor data to detect people, vehicles, structures, and fires in real time, and autonomously tracking targets or avoiding obstacles as a self-flying system.
Classifies people, vehicles, and facilities and extracts position coordinates
Assigns unique IDs to objects and tracks them continuously across frames
One-to-many identification to distinguish insiders from outsiders
Maintains target center coordinates using PID feedback control
Real-time forward obstacle distance sensing and avoidance
Maintains attitude rotation rate and altitude for stable hover control
Live video streaming + bidirectional command transmission
| Main Board | NVIDIA Jetson Orin Nano / Xavier NX |
| Camera Module | CSI or USB camera (supports IMX219, IMX477, etc.) |
| Frame | S500 custom frame (12" or larger) + 3D printed design |
| Flight Controller | Pixhawk / Cube Orange (ArduPilot firmware-based) |
| Servo/Motor | PCA9685 PWM board (Python controllable) |
| Communication | Wi-Fi 5GHz / LTE module (optional) / MAVLink + DroneKit / WebSocket |
| Power | 6S LiPo + Solar Snap (in development) |
Fire:Watch AI
Real-time fire · person · object (identifiable) recognition AI system
FIG.01 — YOLO DETECTION
FIG.02 — SMOKE DETECTION
FIG.03 — EDGE INFERENCE
FIG.04 — FIELD CAMERA
A real-time object recognition AI system based on YOLO. It recognizes not only fire (smoke and flames) but also people and objects (identifiable). Trained on 5,000 smoke images and 300 normal Korean mountain and urban scenes among other diverse datasets. When edge AI devices detect objects from on-site CCTV feeds, they transmit data to the server, which assesses the situation and sends immediate alerts to the control system.
Detects fire, people, and objects in real time from CCTV feeds, outputting bounding boxes and confidence scores
Runs real-time inference on edge devices and immediately transmits detection results to the server
Calculates actual fire status, fire direction, and suppression priority, then alerts the control room
AI auto-classifies clouds, smoke, and false positives to label only actual fires
Real-time video monitoring, detection logs, per-camera settings, and management dashboard
Scheduled retraining of the fire AI to manage false positives and improve accuracy
| Engine | YOLO (You Only Look Once) |
| Training Data | 5,000 smoke images + 300 normal Korean mountain/urban images |
| Detection Classes | Fire / Smoke / Person / Object |
| Inference Environment | NVIDIA Jetson edge device (on-site) + GPU server (dual backend inference) |
| Control System | Web-based real-time monitoring dashboard |