基于惯性传感数据概率密度分布演化特征的分心驾驶状态辨识

Driving Distraction Recognition Based on Probability Distribution Evolution Characteristics of Driving Behaviors

摘要:

不良驾驶行为是道路交通事故的主要致因，近三分之一的事故由分心驾驶引发，辨识分心驾驶状态可有效提升行车安全。然而，当前的分心辨识方法依赖于多源融合感知数据，无法应用于大规模的现存车辆。提出了一种基于泛在惯性测量单元（IMU）数据的两阶段分心辨识方法：第一阶段采用概率密度分布演化，以实现在驾驶行为与行驶工况强耦合情况下的分心状态表征；第二阶段则利用一阶段获得的驾驶行为分布演化特征，采用深度森林算法构建分心驾驶状态辨识模型，以应对复杂的现实场景。为验证所提出的分心辨识方法，使用了上海网约车智能手机惯性测量数据进行实证研究，实验结果表明：实证数据初步验证了所提出方法的有效性，分心驾驶行为特性主要体现在车辆行进方向；与传统表征指标相比，所提出的驾驶行为分布演化特征能够有效提升分心辨识模型性能，准确率、精确率分别提升了20.4 %和10.2 %；所采用的深度森林模型与支持向量机和梯度增强决策树相比，在保持高召回率的同时减少了超10 %的误报情况。

Abstract:

Risky driving behaviors are the main cause of road traffic accidents， with a third of accidents caused by distracted driving. Driving distraction recognition is an efficient approach to improve traffic safety. Current methodologies for driving distraction recognition mainly rely on aggregated multi-sensor data， which limits their extensive application to existing vehicles. Therefore， a two-stage method is proposed in this paper based on inertial measurement unit （IMU） data， a widely available data， for driving distraction recognition. In the first stage， a characterization method based on the evolution of probability density distribution is proposed to represent distracted driving behaviors that are closely coupled with operating conditions. In the second stage， the deep forest algorithm is employed to construct a classification model capable of recognizing driving distraction in complex practical scenarios. An empirical experiment is conducted using IMU data from smartphones in online hailing cars in Shanghai to validate the proposed recognition method. The results indicate that： the distraction recognition method proposed is validated， and the longitudinal characteristics represent the distracted driving behaviors. The proposed characteristics， when compared with the traditional ones， significantly enhance the performance of the model with an increase of 20.4 % in accuracy and 10.2 % in precision. The deep forest model reduces false alarms by more than 10 % while maintaining a high recall rate， compared to support vector machine （SVM） and extreme gradient boosting （XGBoost）.