DNN-Based Prediction of Standard Driving Posture for Vehicle Takeover

Abstract

When drivers need to take over a vehicle during shared autonomy, the standard driving postures based on their body size are the basis of non-driving posture (NDP) motion reconstruction. This study focused on the prediction of standard driving postures using a deep learning neural network (DNN) method. Firstly, the main factors influencing the standard driving posture were extracted through qualitative analysis, and their weights were analyzed using an orthogonal test method. Based on this, the main parameters of the standard driving posture prediction model were determined. Secondly, the point cloud data of typical vehicles on the market were obtained through laser scanning. After extracting the key input and output parameters required for the prediction model through point cloud data processing and feature matching, a dataset of standard driving postures was established. Finally, a supervised learning model using a deep learning neural network (DNN) was established to predict the standard driving postures of different drivers under different vehicle package layouts. This method allows for the quick evaluation of corresponding standard driving postures during non-driving activities, laying the foundation for risk-level assessment of non-driving postures and motion reconstruction in vehicle takeover. The results show that the trained algorithm model can predict standard driving postures with high accuracy and robustness.