Applying Neural Network Methods in Driver Monitoring Systems

Driver monitoring systems are becoming increasingly important and relevant with the growing popularity of ADAS (Advanced Driver Assistance Systems). These systems are designed to address various computer vision tasks with the ultimate goal of enhancing driver safety while operating a vehicle. This study focuses on driver monitoring systems and the use of convolutional neural networks (CNNs) within them. It provides an overview of various classical and contemporary approaches to solving computer vision problems, ranging from traditional machine learning to deep learning and the latest CNN architectures. The structure and fundamental principles of CNNs will be analyzed, along with the most popular and effective CNN architectures used in image classification tasks. The practical part of the study describes a solution for the task of classifying driver hands from frames captured by an in-cabin camera. To address the aforementioned task, the following steps were undertaken: 1) Preparation (annotation) of the DriverMVT dataset. 2) Preprocessing and analysis of the dataset using Python. 3) Examination of classification approaches, fundamental principles of operation, and popular CNN architectures. 4) Implementation and training of models using the PyTorch framework. 5) Analysis of metrics and comparison of models based on specific criteria.