Rased | راصد

Rased is an on-device AI dashcam app built with Flutter. Mounted on the dashboard, it watches the road, detects potholes in real time with a YOLO26n model running locally (LiteRT / TensorFlow Lite), and reports each confirmed pothole once — photo plus GPS — to the municipal backend, without the driver touching the phone.

Documentation:

• docs/live-detection.md — how the live detection pipeline works today: architecture, model, trigger policy, diagnostics, device requirements, what is still unverified on a phone.
• docs/production-roadmap.md — what remains to make it production ready, in order, with exit criteria.
• docs/ai-accuracy-roadmap.md — the ML plan: measuring real accuracy, local data collection, training recipe, edge tuning, and the monthly improvement loop.

---

Key features

• Live detection off the UI thread. All per-frame work (colour conversion, inference, JPEG encoding) runs in a dedicated worker isolate. The preview stays smooth regardless of phone speed. Backend selection is GPU → CPU/XNNPACK (FP16) → CPU, chosen and verified at start-up.
• One report per pothole. A time-based episode policy confirms a pothole over ~300 ms of detections and fires once; spatial dedupe stops repeats on return legs or in traffic.
• Reports never pause detection. Photo from the detection frame, GPS from a continuous stream, uploads through a background queue with offline fallback.
• On-device diagnostics. A 2-second line in the debug log (flutter logs / logcat) shows backend, processed FPS, per-stage timings and UI jank; a per-episode log on the phone supports threshold tuning from real drives. The camera screen itself shows only the driver HUD.
• Manual reporting with local AI pre-check, bilingual (Arabic/English) dark UI, map and admin screens, role-based backend.

Tech stack

• Flutter / Dart (Dart ≥ 3.11)
• tflite_flutter 0.12 (LiteRT), camera (camerax), geolocator, image, path_provider, http, shared_preferences
• Model: YOLO26n exported to .tflite (see YoloModel/)

Getting started

Prerequisites: Flutter SDK (stable), Android SDK / Android Studio, a physical Android phone (camera and GPU are required for anything meaningful).

flutter pub get
flutter analyze
dart test test/frame_sampler_test.dart test/report_trigger_policy_test.dart
flutter build apk --release --split-per-abi   # or: flutter run --profile

Install build/app/outputs/flutter-apk/app-arm64-v8a-release.apk. Always measure with release or profile builds; debug builds are several times slower in the pixel loops.

Windows note: Smart App Control may block Flutter's shader compiler (impellerc.exe). Add an exclusion for the Flutter SDK folder before flutter test, flutter run or flutter build.

Model file

The model ships in assets/ with the naming convention pothole_<arch>_<imgsz>_<precision>[_raw].tflite (currently pothole_yolo26n_640_fp32.tflite). To try another export, add it to pubspec.yaml under flutter: assets: and to TFLiteService.benchModelAssets, then point TFLiteService.defaultModelAsset at it (and set the GPU preference) to bench it; the 2-second diagnostics line in flutter logs shows the effect. Export variants with YoloModel/export_model.py and verify them with assets/check_model.py.

Repository layout

lib/
  main.dart                          app start; pre-warms the AI worker
  screens/live_camera_screen.dart    dashcam screen: frame dispatch, trigger, UI
  services/detection_worker.dart     worker isolate: interpreter, decode, snapshot
  services/frame_sampler.dart        letterbox + rotation + YUV→RGB (pure Dart, tested)
  services/tflite_service.dart       main-isolate facade over the worker (singleton)
  services/report_trigger_policy.dart  when detections become a report (tested)
  services/report_upload_queue.dart  background uploads + offline fallback
  services/episode_log.dart          per-episode JSONL log for tuning
  services/api_service.dart          REST client (auth, hazards)
  widgets/bounding_box_painter.dart  draws boxes on the preview
test/                                pure-Dart unit tests (run with `dart test`)
YoloModel/                           dataset merge, training and export scripts
assets/                              model, classes.txt, check_model.py
docs/                                documentation

Team

• Frontend Engineer: Ahmad Ali
• AI / ML Engineer: Abdallah Abughallous
• Backend Engineer: Abd Alqader Alsa'di

Privacy and permissions

• Camera — to scan the road and capture hazard photos.
• Location (precise) — to attach GPS coordinates to reports.
• Storage (app-private) — report photos awaiting upload, offline queue, diagnostics log.