100% on-device. No cloud. No subscriptions. One APK that auto-tunes from a 2GB budget phone to a flagship with an NPU.
Lyne is a warning-only Advanced Driver Assistance System (ADAS Level 1) that runs entirely on your phone's rear camera and sensors. Mount it on the windshield, start a drive, and it watches the road for you: vehicles closing in, pedestrians stepping out, lane drift, tailgating, and speed-limit signs.
Every model runs on the device. Nothing leaves the phone. It works offline, costs nothing to run, and adapts its own workload to whatever hardware it lands on.
🟢 Real models, not demos. Real YOLOv8n object detection, real drivable-area segmentation, and a real GTSRB traffic-sign classifier are bundled and verified running on a physical device.
| Feature | What it does | |
|---|---|---|
| 🚙 | Forward Collision Warning | Monocular distance plus scale-based time-to-collision, with escalating audio, visual, and haptic alerts |
| 🟩 | Drivable-Area Segmentation | A real-time neural segmentation net paints the road ahead as a green "carpet" and drives lane-keeping |
| 🚶 | Vulnerable Road Users | Detects pedestrians, two-wheelers, cyclists, and auto-rickshaws in the ego path |
| 📏 | Headway Monitor | Continuous following-distance readout in seconds, with a tailgating warning |
| 🛑 | Traffic Sign Recognition | Speed-limit and stop-sign recognition, compared against GPS speed for overspeed alerts |
| 🎯 | Multi-Object Tracking | IoU tracker gives every vehicle a stable ID, so TTC and closing speed are smooth, not jittery |
| 🎚️ | Adaptive Sensitivity | Warning thresholds shift with GPS speed: gentle in the city, earlier on the highway |
| 🗺️ | Trip History + Route Map | Every drive is saved with a route map, stats, and an event timeline |
| 🎥 | Event Dashcam | A rolling ring buffer auto-saves a clip around every critical event, with an in-app player and share |
| 🎛️ | Self-Tuning Tiers | A boot-time probe benchmarks the device and picks input size, model variant, cadence, and backend |
| 🌗 | Auto Day / Night HUD | Ambient-light sensor switches between a daylight and a low-glare night theme |
| 📊 | Drive Sessions | Live trip stats plus an event log exportable as a portable JSON sidecar |
flowchart LR
A[📷 CameraX<br/>rear camera] --> B[Frame Analyzer<br/>downscale + pool<br/>drop on busy]
B --> C{{Inference Scheduler<br/>staggered cadence}}
C -->|every frame| D[🚗 YOLOv8n<br/>objects]
C -->|async thread| E[🟩 TwinLiteNet<br/>drivable area]
C -->|slow cadence| F[🛑 GTSRB<br/>sign classifier]
G[📍 GPS speed] --> H
I[🧭 IMU + light] --> H
D --> H[🔀 Fusion Engine<br/>TTC, headway, LDW, overspeed]
E --> H
F --> H
H --> J[🚨 Alerts<br/>tone + haptics]
H --> K[🖥️ HUD Overlay<br/>gauges + carpet + boxes]
H --> L[🗂️ Event Log<br/>JSON sidecar]
The camera analyzer never blocks: it copies each frame into a reused buffer, closes the image, and submits to a single inference thread with a bounded queue. The heavy segmentation model runs on its own worker thread so it never stalls collision detection. There are zero per-frame allocations in the hot path. Full design notes live in ARCHITECTURE.md.
A capability probe (RAM, cores, GPU, NNAPI, plus a CPU micro-benchmark) classifies the device in under 3 seconds and loads a declarative profile:
| 🅲 Tier C (entry / Go) | |||
|---|---|---|---|
| Input resolution | 384 px | 256 px | 256 px |
| FPS target | 30+ | 15+ | 8+ |
| Object model | YOLOv8n | YOLOv8n nano | YOLOv8n nano |
| Drivable area | Neural seg | Neural seg | Classical seg |
| Backend | NNAPI / GPU | NNAPI / GPU | XNNPACK CPU |
| Pixel format | ARGB_8888 | ARGB_8888 | RGB_565 |
| Sign recognition | On | On | Off |
Thermal status and memory pressure throttle cadence and resolution at runtime, then recover when the device cools down.
| Model | Job | Format |
|---|---|---|
| YOLOv8n (COCO) | Vehicles, people, two-wheelers, stop signs | INT8 TFLite |
| TwinLiteNet | Drivable-area segmentation (the green carpet) | TFLite |
| GTSRB classifier | Speed-limit and stop-sign reading | INT8 TFLite |
| Classical CV | Model-free drivable-area fallback for entry devices | Pure Kotlin |
Conversion scripts and a fine-tuning recipe for India-specific roads (IDD) are in models/.
Just want to try it? Grab the APK from Releases and sideload it.
Build from source:
git clone https://github.com/Sherin-SEF-AI/Lyne-ADAS.git
cd Lyne-ADAS
# point local.properties at your Android SDK, then:
./gradlew :app:assembleDebug
adb install -r app/build/outputs/apk/debug/app-debug.apkRequirements: JDK 17, Android SDK platform 35, build-tools 35. A Gradle wrapper is included.
- minSdk 26, target / compile 35
- Kotlin 2.1, Jetpack Compose, CameraX 1.4, LiteRT 1.0, Play Services Location 21.3
Monocular distance depends on field of view. Lyne reads camera intrinsics from the device when available, and the in-app calibration wizard lets you fine-tune the effective FoV and a distance scale against a known gap. Values persist per device.
- India Driving Dataset (IDD) fine-tuned segmentation weights
- Dashcam event clip recording (ring buffer)
- Cyclist and animal classes for rural roads
- Bluetooth auto-start on drive detection
- CI build and instrumented tests
🚨 Lyne is a warning-only aid. It never controls the vehicle. Keep your eyes on the road and your hands on the wheel.
- Monocular distance and TTC are estimates from bounding-box geometry on a flat-road assumption. Low-confidence estimates are suppressed, never escalated to a hard alert.
- Drivable-area and sign recognition accuracy depend on lighting, weather, and marking quality.
- HUD overlay alignment is approximate (it assumes a center-crop preview).
- This is research and educational software, provided as is, with no warranty.
sherin joseph roy ✉️ sherin.joseph2217@gmail.com 🔗 https://github.com/Sherin-SEF-AI
Released under the MIT License.