Abstract: Recognizing the current driving behavior patterns of vehicles accurately is an urgent technical problem to be solved in the field of autonomous driving. To analyze driving behavior patterns accurately and improve the accuracy and reliability of the recognition model, the multi-source parameter information such as the driving behavior data and visual characteristic data of 20 drivers under high-speed conditions was collected through the naturalistic driving test, and four typical driving behavior patterns (free driving, car following, left and right lane changes) and multi-source parameter coupling characteristics were analyzed. Four types of driving behavior patterns index sets were determined based on the principal component method, random forest decision tree and support vector machine were used to establish recognition models. The recognition results were compared after model training and learning, the model with better recognition effect was furtherly optimized, and the temporal variation characteristics of the recognition accuracy of the optimized model for four types of driving behavior patterns were analyzed. The results show that the total recognition rates of the support vector machine model, the random forest decision tree model, and the random forest optimization model based on the multi-layer perceptron neural network are 89.4%, 90.5% and 91.9%, respectively. AUC (area under the curve) values of the optimized model are all greater than 0.93 under four driving behavior patterns, which can better recognize the current driving behavior patterns. In addition, the recognition accuracy of the random forest optimization model for four types of driving patterns shows a trend of first increasing and then gradually becoming stable over time, and the recognition accuracies of the car following and free driving patterns at the same time are greater than that of left and right lane changes. The results will provide technical support and theoretical basis for the decision-making and control strategy formulation of high-level auto drive system.

Key words: traffic engineering; driving patterns; artificial intelligence; high-speed conditions; naturalistic driving test; recognition model

摘要： 准确识别车辆当前驾驶行为模式是自动驾驶领域亟待解决的技术问题。为实现驾驶行为模式精准解析，提高模型识别精度和可靠性，通过开展自然驾驶试验，采集高速工况下20名驾驶人的驾驶行为数据及视觉特性数据等多源参数信息，分析4类典型驾驶行为模式（自由行驶、跟车、左换道、右换道）运行规律及多源参数耦合特性。基于主成分分析法确定4类驾驶行为模式表征指标集，使用支持向量机、随机森林决策树算法建立驾驶行为模式识别模型，通过学习训练，分析比较模型识别结果，对识别效果较好的模型进一步优化，分析优化模型对4类驾驶行为模式识别精度的时序性变化特征。研究结果表明：支持向量机模型、随机森林决策树模型、基于多层感知器神经网络的随机森林优化模型总体识别精度分别为89.4%、90.5%、91.9%；4类驾驶行为模式的AUC (area under the curve) 值均高于0.93，可较好地识别车辆当前驾驶行为模式。此外，随机森林优化模型对4类驾驶行为模式的识别精度均随时间推移，呈现先增长后趋于稳定的变化态势，且同一时刻的自由行驶及跟车模式识别精度高于向左及向右换道模式。研究结果可为高级别自动驾驶系统决策及控制策略的制定提供理论基础和技术支持。

关键词: 交通工程；驾驶行为模式；人工智能；高速工况；自然驾驶试验；识别模型