Car Nebula ADAS — On-Device TFLite Models

Two TFLite models used by Car Nebula for real-time Advanced Driver Assistance (ADAS) running entirely on-device on Android automotive hardware.

---

Models

1. lane_detector.tflite — Lane Detection

Property	Value
Base model	Ultra-Fast Lane Detection (cfzd/Ultra-Fast-Lane-Detection)
Architecture	ResNet-18 backbone
Dataset	TUSimple (highway lanes, US dashcam footage)
Original weights	tusimple_res18.pth (official pre-trained)
Input shape	[1, 288, 800, 3] — float32 or int8, NHWC
Input normalization	ImageNet: mean [0.485, 0.456, 0.406], std [0.229, 0.224, 0.225]
Output shape	[1, 201, 56, 4] or [1, 4, 56, 201]
Output format	201 grid bins (200 x-positions + 1 no-lane) × 56 row anchors × 4 lanes
License	MIT

Conversion pipeline:

tusimple_res18.pth  →  ONNX (opset 11)  →  TF SavedModel  →  TFLite

Conversion script: convert_ufld.py (included in the Car Nebula Android repo).

Row anchors (TUSimple, 56 rows, pixel Y in 288-px input):

64 68 72 76 80 84 88 92 96 100 104 108 112 116 120 124 128 132 136 140
144 148 152 156 160 164 168 172 176 180 184 188 192 196 200 204 208 212
216 220 224 228 232 236 240 244 248 252 256 260 264 268 272 276 280 284

---

2. efficientdet_lite0.tflite — Object Detection

Property	Value
Model family	EfficientDet Lite0
Source	TensorFlow Hub
Input shape	[1, 320, 320, 3] — uint8, NHWC
Output	Bounding boxes · class scores · class labels · detection count
Classes	90 COCO classes
Score threshold	0.38 (used by Car Nebula pipeline)
License	Apache 2.0

---

How the pipeline uses both models

Camera frame (CameraX / USB UVC)
        │
        ├──► EfficientDet Lite0 (every 3 frames)
        │         └──► AdasBox list: label, bounding rect, estimated distance
        │                   └──► HUD overlay: boxes, collision warning, top-down view
        │
        └──► UFLD ResNet-18 (every 3 frames)
                  └──► Lane boundary points (left/right, top/bottom)
                            └──► Camera overlay: seg mask, lane lines
                                 HUD: road corridor, departure warning

Results are temporally smoothed between inference runs so the UI always has something to render even on frames that skip inference.

---

Usage (Android / TFLite Java API)

// Load from downloaded file
val model = FileInputStream(file).channel.use { ch ->
    ch.map(FileChannel.MapMode.READ_ONLY, 0, file.length())
}
val interpreter = InterpreterApi.create(
    model,
    InterpreterApi.Options()
        .setRuntime(InterpreterApi.Options.TfLiteRuntime.FROM_SYSTEM_ONLY)
        .setNumThreads(2)
)

---

Licenses

Model	License
lane_detector.tflite (UFLD ResNet-18)	MIT
efficientdet_lite0.tflite (TFHub)	Apache 2.0