robonix.service.fire_detect

service v0.1.0

Drone fire & smoke detection service -- pulls MJPEG frames from the DJI bridge APK, runs YOLO to detect flame/smoke; attaches the current GPS coordinates on a hit

View source Open README on GitHub JSON
README service-fire-detect-rbnx@main

fire_detect -- drone fire & smoke detection primitive

A RoboNIX primitive package independent of dji_msdk, in the robonix/service/fire_detect/* namespace. It coexists with dji_msdk (robonix/primitive/drone/*) and can be called by the executor simultaneously:

# Drone takeoff / state
rbnx call robonix/primitive/drone/takeoff '{"altitude": 5.0}'
rbnx call robonix/primitive/drone/state

# Fire & smoke detection (single frame: grab one frame and infer; returns GPS on a hit)
rbnx call robonix/service/fire_detect/fire_detect '{"min_confidence": 0.4}'

# Real-time fire & smoke monitoring (background continuous frame inference, polled by an external loop)
rbnx call robonix/service/fire_detect/start_monitor
rbnx call robonix/service/fire_detect/monitor_state      # poll repeatedly to read the latest alert/gps
rbnx call robonix/service/fire_detect/stop_monitor

Architecture

RoboNIX executor
   ├─ robonix/primitive/drone/*   → dji_msdk (drone control)
   └─ robonix/service/fire_detect/*  → fire_detect (fire & smoke detection)
          │ MJPEG stream pull + YOLO inference
          ▼
   DJI bridge APK :8080/api/video + /api/capture_gps
          │
          ▼
   drone camera frame

Fire & smoke detection reads the stream and does not control the aircraft; GPS is fetched live via /api/capture_gps and only attached on a hit, for use by alert primitives / upper layers.

Layout

fire_detect/
  package_manifest.yaml            # package metadata + capabilities + config
  capabilities/                    # contracts (Schema A, consumed by rbnx codegen --mcp)
  fire_detect/
    __init__.py
    backend.py                     # frame grab + inference + GPS (SDK-agnostic, pure Python)
    driver.py                      # @vision.mcp(...) dispatch layer
    main.py                        # standalone REPL (framework-free debugging)
  models/                          # fire & smoke YOLO weights (default fallback ultralytics yolov8n)
  scripts/build.sh / start.sh
  requirements.txt

Model

Model loading priority in backend.py:

  1. weights specified by config.model_path (.pt / .engine / .onnx)
  2. the FIRE_MODEL_PATH environment variable
  3. the first weights file in the models/ directory
  4. fallback to the official ultralytics yolov8n (no fire/smoke-specific classes; only validates the pipeline)

Production tip: use a self-trained fire/smoke two-class YOLO weights file; export a TensorRT .engine on Jetson (FP16) for real-time inference.

Dependencies

  • requests (stream pull / GPS)
  • opencv-python (MJPEG decoding)
  • ultralytics (YOLO inference; for TensorRT deployment only torch + engine are needed)