Road Accident Prediction Using Machine and Deep Learning Techniques

Abstract

This project aims to deal with the critical issue of drowsiness detection in drivers using convolutional neural networks (CNN) with machine learning. The study involves extensive preprocessing of a diverse dataset comprising images depicting yawning and non-yawning faces, along with images of open and closed eyes. Subsequently, different machine learning techniques, such as Logistic Regression, Support Vector Machine, AdaBoost Classifier, Decision Tree Classifier, and XG Boost Classifier, are trained on augmented data to classify drowsiness states based on facial and eye features. Additionally, a sophisticated CNN architecture is suggested and trained to enhance classification accuracy further. The efficiency of every model is rigorously assessed utilizing a number of metrics, such as F1-score, recall, accuracy, and precision. The findings of the trial show how much better the CNN model is compared to conventional machine-learning techniques with regard to accuracy and robustness in drowsiness detection. Further research could focus on exploring advanced data augmentation techniques and real-time implementation for practical deployment in vehicles.

Furthermore, to model training and evaluation, this project also emphasizes the significance of interpretability and explainability in drowsiness detection systems. Understanding the decision-making process of AI models is crucial, especially in safety-critical applications like driver monitoring. Hence, alongside achieving high accuracy, the focus is on interpreting the CNN model's predictions and identifying the key features contributing to drowsiness detection. Techniques such as saliency mapping and feature visualization are employed to elucidate the model's decision boundaries and highlight relevant regions in input images. By enhancing interpretability, drivers where anyone with an interest can acquire knowledge of the factors influencing drowsiness detection, thereby fostering trust and facilitating informed decision-making. Future research could concentrate more intently on interpretability techniques and develop user-friendly interfaces to present model insights effectively, thus promoting the adoption and acceptance of drowsiness detection systems within the automotive industry.