This model is a high-performance Ocular State Classifier developed as part of the Vigilance AI suite. It is designed to distinguish between "Open" and "Closed" eyes in real-time to detect driver drowsiness and microsleep events. By leveraging the MobileNetV2 architecture, the model achieves industrial-grade accuracy while maintaining the ultra-low latency required for automotive safety applications. Developer: Abdullah, Salman Shah, Sheheryar Ilyas Task: Binary Image Classification (Ocular State) Base Architecture: MobileNetV2 (Transfer Learning) Classes: Closed, Open Validation Accuracy: ~90%+

The model utilizes a MobileNetV2 backbone, chosen for its Depth-wise Separable Convolutions which significantly reduce computational load without sacrificing depth. Input Shape: 224 x 224 x 3 (RGB) Freezing Strategy: The first 100 layers were frozen to utilize pre-trained ImageNet weights for general feature extraction. Custom Head: Global Average Pooling: Redimensionalization of spatial features. Dense Layer: 128 neurons with ReLU activation. Dropout: 0.5 (50%) to ensure robustness against overfitting. Output Layer: Softmax/Dense layer for binary probability distribution.

Dataset: Merged high-variance dataset including the MRL Eye Dataset, consisting of thousands of samples featuring various ethnicities, lighting conditions, and ocular occlusions (e.g., glasses). Optimization: Adam Optimizer with a phased learning rate approach (starting at 1e^−4 and fine-tuning at 1e^−5). Data Augmentation: To simulate real-world driving shifts, images were randomly rotated, zoomed, and horizontally flipped during training.

Metric Value Accuracy 90.27% Inference Time < 10ms (CPU) Framework TensorFlow / Keras

Drowsiness Detection Systems: The core component for triggering "Wake Up" alarms. Human-Computer Interaction (HCI): Gaze tracking and eye-state triggers. Fatigue Research: Analyzing blink duration and ocular patterns in high-stress environments.

You can implement this model in Python using TensorFlow:

import tensorflow as tf import numpy as np

model = tf.keras.models.load_model("eye_classifier.h5")

def check_eye(image_crop): # Preprocess to match training specs img = cv2.resize(image_crop, (224, 224)) img = img.astype("float32") / 255.0 img = np.expand_dims(img, axis=0)

prediction = model.predict(img)
state = "Closed" if np.argmax(prediction) == 0 else "Open"
return state

Occlusions: Very dark or reflective sunglasses may hinder the accuracy of the underlying landmark detector used to crop the eye region. Extreme Darkness: Standard RGB cameras may struggle in pitch-black cabins; performance is optimal when paired with an IR-sensitive camera or cabin lighting.

This model is a pillar of the Vigilance AI project, which combines deep learning classification with geometric EAR (Eye Aspect Ratio) math to provide a redundant, fail-safe safety shield for modern drivers.