Driving distraction is a major cause of frequent traffic accidents. This research endeavors to construct a high-precision model for recognizing distracted driving behaviors by leveraging driver posture characteristics, with the overarching objective of augmenting traffic safety measures. The OpenPose project is used to extract key points, and a detection method integrating Principal Component Analysis (PCA) and XGBoost is proposed. PCA reduces the dimensionality of the driver’s upper body key point distance and joint angle information to simplify the model while retaining essential information. XGBoost is then applied to discriminate and detect ten categories of distracted driving behaviors. Taking into account driver pose information, the proposed method demonstrates high precision in detecting ten distinct driving behaviors, achieving an overall accuracy of 95.8% on the publicly available State Farm dataset and 94.6% on a self-constructed dataset. Compared to deep learning methods, this approach is less sensitive to lighting and minimally affected by environmental conditions as it extracts pose feature information based on key points. It also reduces model parameters and has streamlined training procedures. However, there are challenges in misclassifying some similar driving behaviors. The pose estimation-based method shows excellent performance in practical applications, and future studies plan to incorporate facial and hand information to further improve detection accuracy.
Lv et al. (Thu,) studied this question.