基于序贯稀疏自编码器和高斯混合模型的驾驶行为分析

Driving Behavior Analysis Using Sequential Sparse Auto-encoder and Gaussian Mixture Model

基于驾驶数据,驾驶行为分析方法能够获得隐藏的驾驶行为信息,进而实现驾驶风格识别等应用。随着传感器技术的发展,先进驾驶辅助系统需分析的驾驶数据的规模和维度不断增加,这提升了驾驶行为分析结果的有效性和普适性,但也给数据分析工作带来了挑战。因此,准确高效的驾驶行为分析方法对于先进驾驶辅助系统的作用越发重要。针对大规模、高维驾驶数据集,本文提出了一种基于序贯稀疏自编码器和高斯混合模型的驾驶行为分析方法。首先,为了有效提取驾驶数据的低维特征,该方法改进了稀疏自编码器在预训练阶段的损失函数,降低了模型参数易落到局部最优的风险;然后,该方法基于线性映射将提取到的驾驶特征映射到颜色空间,实现了驾驶行为的可视化;最后,该方法使用高斯混合模型对提取到的驾驶特征进行聚类,实现了驾驶风格的识别。真实驾驶数据的验证结果表明,所提算法可以提取到比传统算法更有区分度的驾驶特征,并在轮廓系数等性能指标下都取得了更好的驾驶风格识别效果。

Based on driving data,driving behavior analysis methods can extract hidden driving behavior information and enable applications such as driving style recognition.With the development of sensor technology,the scale and dimensionality of driving data required by advanced driver assistance systems are constantly increasing,which poses great challenges for driving behavior analysis.However,this also poses challenges for data analysis.Therefore,efficient and accurate driving behavior analysis methods are becoming increasingly important for advanced driver assistance systems.A driving behavior analysis method based on sequential sparse auto-encoder and Gaussian mixture model was proposed for large-scale and highdimensional driving data sets.First,the loss function of the sparse auto-encoder was modified to effectively extract the low-dimensional representations of driving data.Then,driving behavior was visualized by projecting the extracted features into the color space using linear mapping.Finally,driving style recognition was performed by clustering the extracted features using a Gaussian mixture model.The experiments on real driving data show that the proposed method can effectively extract differentiated driving features and outperform other methods in driving style recognition according to indicators such as silhouette coefficients,achieving efficient and accurate driving behavior analysis.