🚗 AlertDrive AI - Driver Drowsiness & Safety Telemetry Model
AlertDrive AI is a lightweight, production-grade driver drowsiness detection system designed for high-performance deployment on both cloud servers and resource-constrained edge devices (e.g., in-vehicle embedded systems, Raspberry Pi, and mobile devices).
This repository hosts both the fully optimized fine-tuned functional Keras Model (.h5) and the highly optimized Float16 Quantized TFLite Model (.tflite).
📊 Model Architecture & Performance
- Base Architecture: MobileNetV2 (Unfrozen from Layer 100 onwards for deep task specialization)
- Input Dimensions: 224x224x3 RGB
- Classification Target: Binary (0 =
DROWSY, 1 =NATURAL/ Alert)
📈 Benchmarks
| Model Variant | Test Accuracy | ROC AUC | Model Size | Target Deployment Platform |
|---|---|---|---|---|
| Keras Fine-Tuned (.h5) | 93.18% | 0.9890 | 13.31 MB | High-end Edge Systems / Cloud Servers |
| Quantized TFLite (Float16) | 93.32% | 0.9890 | 4.94 MB | Microcontrollers / Low-resource Edge Devices |
The Float16 quantized model achieves a 62.9% reduction in model footprint while perfectly preserving evaluation accuracy.
🛠️ Usage & Integration
1. Integrating the High-Accuracy Keras Model (.h5)
from huggingface_hub import hf_hub_download
import tensorflow as tf
# Download and load model
model_path = hf_hub_download(
repo_id="btehubsolutions/alertdrive-model",
filename="alertdrive_finetuned_model.h5"
)
model = tf.keras.models.load_model(model_path)
# Perform inference
# input_image must be preprocessed to (224, 224, 3) and rescaled by 1/255.0
prediction = model.predict(input_image)
class_label = "NATURAL" if prediction[0][0] > 0.5 else "DROWSY"
2. Integrating the Ultra-Lightweight Quantized TFLite Model (.tflite)
import numpy as np
import tensorflow as tf
from huggingface_hub import hf_hub_download
# Download TFLite model file
tflite_path = hf_hub_download(
repo_id="btehubsolutions/alertdrive-model",
filename="alertdrive_model_quantized.tflite"
)
# Initialize Interpreter
interpreter = tf.lite.Interpreter(model_path=tflite_path)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
# Input image array shape must be [1, 224, 224, 3] and float32 normalized
interpreter.set_tensor(input_details[0]['index'], input_image_array)
interpreter.invoke()
output_data = interpreter.get_tensor(output_details[0]['index'])
class_label = "NATURAL" if output_data[0][0] > 0.5 else "DROWSY"
🚗 Production-Grade Face Telemetry & Fallback Pipeline
To complement the classifiers, the production deployment codebase supports active facial telemetry extraction:
- Eye Aspect Ratio (EAR): Computes eye closure to monitor blink patterns and detect microsleep.
- Mouth Aspect Ratio (MAR): Monitors facial configurations to track yawning.
- Head Pose Estimation (Pitch/Yaw/Roll): Identifies side-glancing or active distraction behaviors.
- PERCLOS Safety Scoring: Tracks eye closure percentage over a 60-frame rolling window to log warning levels (
SAFE,WARNING ALERT,CRITICAL ALERT).
Mathematical Fallback Engine
To ensure driver-safety features never experience fatal runtime crashes, the pipeline incorporates an automatic Mathematical Fallback Pipeline that gracefully estimates nominal metrics even if local environments experience library incompatibilities or missing system camera devices.
👥 Developers
Developed by the BTEHub Team for deployment within Nigerian Road Safety and transport environments.