作者：Yuanzhuo Xu, Pengfei Sun, Liwei Jing, Shaowu Wu, Kejiang Xiao（肖克江）, Xiaoguang Niu*

出版刊物：Knowledge-Based Systems

出版时间：2026年

内容摘要：

Perceiving drivers' status is highly valuable to auto services and insurance companies for driver assessments and improvements in driving safety. However, most existing status perception methods face challenges in deployment, privacy protection, and data collection to extract distinctive features from driving data. In this paper, we propose InvisiSense, a novel privacy-preserving driving status perception system that uniquely integrates stochastic cloned Extended Kalman Filter (EKF) with a deep residual network to achieve high-accuracy motion estimation and driving status classification. Unlike existing methods relying on visual data, InvisiSense only uses IMU data from smartphones, ensuring better privacy protection and easier deployment. A classifier based on the attention mechanism is built for better performance on status perception. Experiment results on invisible driving IMU data over 3268.1 km of trips show that the proposed system can achieve the same level of effectiveness compared to the existing method from different data sources (e.g., visual) and the average perception accuracy of different status can reach 72.3%.

感知驾驶员状态对于汽车服务机构和保险公司而言具有重要价值，可用于驾驶员评估及驾驶安全改进。然而，现有大多数状态感知方法在部署、隐私保护以及数据采集以提取驾驶数据中具有区分度特征方面面临挑战。本文提出InvisiSense，一种新型隐私保护驾驶状态感知系统，其独特地将随机克隆扩展卡尔曼滤波（EKF）与深度残差网络相结合，以实现高精度运动估计与驾驶状态分类。与依赖视觉数据的现有方法不同，InvisiSense仅使用来自智能手机的IMU数据，从而确保更好的隐私保护与更便捷的部署。为实现更优的状态感知性能，构建了基于注意力机制的分类器。在覆盖3268.1公里行程的不可见驾驶IMU数据上的实验结果表明，所提系统能够达到与来自不同数据源（如视觉）的现有方法相当的效果，且不同状态的平均感知准确率可达72.3%。