Analysis of Hybrid Feature and Data Adaptation Techniques for User-Dependent and User-Independent Driver Condition Recognition

Distracted drivers are one cause of many traffic accidents. Therefore, there is a need for a system that can provide warnings to distracted drivers. Many accurate systems have been reported in previous studies, but some systems cannot recognize new users. This research seeks to develop an accurate and robust system to detect driver distraction using the Convolutional Neural Network (CNN) model for these conditions to improve the robustness of new drivers. This research proposes hybrid features and retraining with adapted data. Hybrid features combine natural features from RGB color composition with artificial features from Scale Invariant Feature Transform (SIFT). This is called SIBlend. For reference, we use RGB features only. The model, named RGBumble, achieves the best accuracy, but it is not able to recognize new drivers. While SIBlend has lower accuracy, it is more robust when dealing with new drivers. The results of RGBumble mode reaches $\mathbf{98. 50 \%}$ and SIBlend $\mathbf{99. 00 \%}$. The results of RGBumble for new users are the best at $\mathbf{90. 00 \%}$ and SIBlend at 90.4%.