drone-traffic-monitor

Real-time vehicle detection and tracking from a DJI Mini 4 Pro drone, with on-device YOLO inference, live web dashboard, and a post-flight processing pipeline for building traffic datasets.

Built as a pedagogical tool for the Economics degree program (Licenciatura en Economia) at Universidad de Buenos Aires (UBA), to teach students about real-world data collection, computer vision pipelines, and statistical dataset construction.

1. On-device detection: YOLOv8 runs directly on a Samsung S23+ connected to the drone via DJI MSDK V5. Vehicles are detected, tracked, and classified in real time.
2. Vehicle tracking: A centroid-based tracker deduplicates vehicles across frames, ensuring each vehicle is counted once even when the drone is stationary.
3. Live dashboard: A web dashboard (served from the phone) shows telemetry, detection stats, and a 7x7 grid of recent vehicle crops in real time.
4. Data persistence: Every unique vehicle is logged to SQLite with class, color, confidence, bounding box, GPS coordinates, altitude, gimbal angle, and a cropped image.
5. Post-flight pipeline: A Python pipeline re-processes the 4K video from the drone's SD card, producing high-resolution crops, embedding-based deduplication, and a clean dataset.

DJI Mini 4 Pro (4K camera)
    |
    | OcuSync / DJI O4
    v
DJI RC-NS (remote controller)
    |
    | USB-C
    v
Samsung S23+ (Kotlin app)
    +-- MSDK V5: video feed, telemetry, camera control
    +-- YOLOv8 (ONNX Runtime): vehicle detection
    +-- Vehicle tracker: centroid-based dedup
    +-- SQLite: detection + telemetry persistence
    +-- NanoHTTPD: live web dashboard (:8080)
    +-- Crop storage: JPG per vehicle
    |
    | WiFi (same network or Tailscale)
    v
Notebook browser: http://<phone-ip>:8080
    +-- Real-time telemetry
    +-- Detection feed (7x7 crop grid)
    +-- Flight stats

Post-flight:
    4K video (SD card) + detections.db
        |
        v
    Python pipeline (process_flight.py)
        +-- YOLOv8s/m on 4K frames
        +-- ResNet18 embeddings for dedup
        +-- Color classification
        +-- ROI filtering
        +-- Optional SAHI for small objects
        |
        v
    Clean dataset: HD crops + CSV + SQLite

python pipeline/process_flight.py \
    --video DJI_0001.MP4 \
    --db detections.db \
    --output results/ \
    --model yolov8s \
    --roi-top 0.15 --roi-bottom 0.80

The DJI RC-N2 also works. The DJI RC 2 (with built-in screen) does not support MSDK apps.

• JDK 17
• Android SDK (platform 35, build-tools 35.0.0, NDK 21.4.7075529)
• DJI Developer account and App Key
• Python 3.10+ with ultralytics, opencv-python, torch, torchvision
• ADB and scrcpy for development

cd app/

# Copy and edit configuration
cp gradle.properties.example gradle.properties
cp local.properties.example local.properties
# Edit both files with your paths and DJI App Key

# Download YOLO models to assets/
cd app/src/main/assets/
python3 -c "
from ultralytics import YOLO
YOLO('yolov8n.pt').export(format='onnx', imgsz=640, simplify=True)
# For VisDrone model:
import urllib.request
urllib.request.urlretrieve(
    'https://huggingface.co/mshamrai/yolov8n-visdrone/resolve/main/best.pt',
    'yolov8n_visdrone.pt'
)
YOLO('yolov8n_visdrone.pt').export(format='onnx', imgsz=640, simplify=True)
"
cd ../../../..

# Build
./gradlew assembleDebug

# Install (via ADB over WiFi)
adb install -r app/build/outputs/apk/debug/app-debug.apk

1. Power on the DJI Mini 4 Pro and RC-NS
2. Connect phone to RC-NS via USB-C
3. Grant USB permission when prompted
4. The app initializes the SDK, connects to the drone, and starts the video feed

From any device on the same network:

http://<phone-ip>:8080

Or via ADB port forwarding:

adb forward tcp:8080 tcp:8080
# then open http://localhost:8080

If you're building your own DJI MSDK V5 app, here are the non-obvious issues we encountered:

1. Helper.install(this) must be called in Application.attachBaseContext(), not onCreate(). Without it, all dji.v5.manager.* classes throw NoClassDefFoundError.

2. Don't reference DJI classes in the Application class body. The classloader resolves them before attachBaseContext runs. Keep DJI imports in Activity classes only.

3. Use cameraStreamManager, not videoStreamManager. The latter is deprecated and returns null for newer drones like Mini 4 Pro.

4. AvailableCameraUpdatedListener has two abstract methods. Missing onCameraStreamEnableUpdate causes AbstractMethodError at runtime (no compile-time warning).

5. The video stream is always 1080p. No API can change this — it's a bandwidth limitation of the RC-to-aircraft link. 4K is only available on the SD card.

6. ComponentIndexType.LEFT_OR_MAIN is the correct camera index for Mini 4 Pro.

The DJI Mini 4 Pro is a sub-249g drone classified as a "very light" UAV. In Argentina (ANAC regulations), this class can be operated in urban areas without special permits, provided the operator maintains visual line of sight and stays away from airports and restricted zones.

MIT

• DJI Mobile SDK V5
• Ultralytics YOLOv8
• ONNX Runtime
• NanoHTTPD
• VisDrone Dataset