为弥补传统驾驶风格识别方法对横向交互行为刻画不足的问题，本文融合车辆换道行为特征优化识别模型，通过构建多维度特征体系并结合无监督聚类方法，实现兼顾纵向运动与横向交互的驾驶风格识别。以“交通之眼”数据集SQM2子集为研究对象，基于车辆轨迹与配套视频数据，构建道路二维坐标系，提取平均速度、速度标准差、平均加速度、加速度标准差、冲击度平均值、冲击度标准差等6项传统纵向特征，以及换道频率、换道时长、换道起始速度、换道起始加速度、换道间距等5项换道行为特征。对11维特征进行Z-score标准化，并采用主成分分析（Principal Component Analysis, PCA）降维和K-means聚类识别驾驶风格。结果表明：在兼顾样本覆盖率、无噪声样本比例和结果可解释性的前提下，K=3时能形成较稳定的聚类结构；与仅采用传统特征的聚类结果相比，融合换道特征后1 420个有效车辆样本中有149个样本的标签发生变化，占比10.49%，一致率为89.51%；3类样本在速度水平、加速度波动性、换道频率和换道间距的相关指标上表现出明显差异，其中平均速度、加速度标准差和换道频率对类别区分作用较为突出。研究表明，换道行为特征能为传统纵向运动信息提供补充，为边界样本提供额外的判别依据；所得的激进型、一般型和保守型标签属于基于样本分布的相对分类结果，可为驾驶行为建模与个性化交通管理提供数据参考。

To address the limitations of traditional driving style identification methods in characterizing lateral interactive behaviors, this study optimizes the recognition model by integrating lane-changing behavior characteristics. By constructing a multidimensional feature system and applying unsupervised clustering, this paper achieves driving style identification that considers both longitudinal motion and lateral interactions. This study takes the SQM2 subset of the TrafficEye dataset as the research object. Based on vehicle trajectories and corresponding video data, a two-dimensional road coordinate system is constructed to extract six traditional longitudinal features: mean speed, speed standard deviation, mean acceleration, acceleration standard deviation, mean jerk, and jerk standard deviation; along with five lane-changing behavior features: lane-changing frequency, lane-changing duration, lane-changing initiation speed, lane-changing initiation acceleration, and lane-changing gap. The 11-dimensional features are standardized using Z-score normalization, and driving styles are identified through Principal Component Analysis (PCA) for dimensionality reduction, followed by K-means clustering. The results show that, considering sample coverage, the proportion of noise-free samples, and result interpretability, a stable clustering structure is formed when K=3. Compared with clustering using only traditional features, the incorporation of lane-changing features results in label changes for 149 out of 1 420 valid vehicle samples, accounting for 10.49%, with a consistency rate of 89.51%. The three clusters exhibit significant differences in terms of speed level, acceleration variability, lane-changing frequency, and lane-changing gap, among which mean speed, acceleration standard deviation, and lane-changing frequency contribute notably to cluster differentiation. The study demonstrates that lane-changing behavior features can complement traditional longitudinal motion information and provide an additional discrimination basis for boundary samples. The resulting aggressive-type, normal-type, and conservative-type labels represent relative classifications based on sample distribution, which can serve as a data reference for driving behavior modeling and personalized traffic management.