【摘要】 为提高智慧交管建设水平，提升执法效率、降低执法成本，对交通事故中当事人责任智能快速划分进行研究。使用Pearson相关系数来计算全部特征与事故责任的相关系数，挑选出与事故责任划分高度相关的数据特征；基于道路交通事故数据及挑选出与事故责任划分明显相关的10个因素为评价指标，使用高效梯度提升决策树(XGBoost)算法对事故责任进行建模预测，结果相对准确，为78.9%,但存在模型对缺失样本分裂方向的处理能力有限及模型过拟合问题；通过参数优化和模型融合方法对XGBoost算法进行优化。结果表明：优化后的算法能有效地自动学习出缺失样本的分裂方向，模型融合对缺失值的处理率已提升至94.8%,处理率提升了50%,缺失样本的分裂方向通过模型融合基本全部得到有效学习，预测结果对比原始算法准确度提升至87.2%,交叉验证结果也表明该算法在交通事故责任智能认定中的适用性。更多还原

【Abstract】 To improve the construction level of intelligent traffic management, the efficiency of law enforcement, reduce the cost of law enforcement as well, and provide a reference for the traffic management department to deal with the accident fairly, the intelligent rapid division of the responsibility of the parties in traffic accidents is studied in this paper. Firstly, the Pearson correlation coefficient is used to calculate the correlation coefficient between all characteristics and accident liability, and 10 data characteristics(including accident liability itself); Secondly, 10 factors that are related to the division of accident responsibility are selected as the evaluation indexes by the correlation coefficient, and the high-efficiency gradient promotion decision tree algorithm(XGBoost) is used to model and predict the accident responsibility based on the road traffic accident data, with the relative accuracy of 78.9%. However, the problems such as the limited ability of the model to deal with the splitting direction of the missing samples and the overfitting of the model exist. Finally, the XGBoost algorithm is optimized by parameter optimization and model fusion, which solves the problem of the limited learning ability of the original XGBoost algorithm in the direction of splitting a large number of missing values. Compared with the original XGBoost algorithm, the comprehensive optimization, and verification of the road traffic accident responsibility fast identification model are completed. The results show the optimized algorithm can automatically learn the splitting direction of missing samples, and the fusion classification model is used to fit all data sets. It is found that the processing rate of missing values has been improved to 94.8%, and the processing rate has been increased by 50%. The splitting direction of missing samples has been almost effectively learned through model fusion. Compared with the original algorithm, the accuracy of the prediction results has been improved to 87.2%. The cross-validation results also show the applicability of the algorithm in the intelligent identification of traffic accident responsibility.更多还原