Abstract: To thoroughly disclose the relationship between taxi speeding behaviors and road characteristics at segmentlevel, this study extracted large-scale taxi speeding behaviors from GPS trajectories collected in downtown Chengdu. The characteristics of speeding behaviors consist of speeding linear density (SLD) and speeding severity (SR) are sorted in different percentiles. Ten types of road characteristics were selected as underlying contributing factors. To restrict the interference of spatial effects, this study examined the spatial correlation among the characteristics of speeding behaviors and developed three types of spatial models. The models include spatial intrinsic conditional autoregressive model (ICAR), spatial error model (SEM), and spatial lag model (SLM), which are based on the Lognormal prior to the response variables (e.g., SLD and SR). In addition to the spatial correlation, the models were extended to be incorporated with random parameters to capture the unobserved heterogeneity among roadways. The results indicate that the spatial models outperform the traditional model with better goodness-of-fit since all the speeding characteristics are observed with severe spatial correlation. We also found the performance of a certain model varies across the type of response variables. In detail, random parameter ICAR model outperforms others on modeling SLD, while SRs on various percentiles are best fitted by different spatial models. The factors of speed limit, roadwaycross-section, and non-motorized vehicle lane have heterogenous effects on the speeding characteristics. The estimates also indicate that the speed limit, roads without divider, non-motorized vehicle lane, and work zone are significantly associated with SLD. Speed limit, non-motorized vehicle lane, overpass or road tunnel, work zone, and road length are significantly related to SR.

Key words: traffic engineering, road characteristics, group-level random parameter spatial model, speeding behavior; intrinsic conditional autoregressive model, spatial simultaneous autoregressive mode

摘要： 为准确揭示道路因素与路段出租车超速行为特征的关联关系，本文以成都市中心城区主 干道路段为研究单元，通过车载GPS设备获取大范围出租车超速行为样本。在筛选出10类典型城市道路特征的基础上，选择路段超速线密度和不同百分位数超速严重度作为研究对象；为减少空间效应对估计的干扰，本文验证了超速特征的空间相关性，并分别构建基于对数-高斯分布的标准线性模型和空间条件自回归模型、空间误差模型及空间滞后模型这3类空间模型，探究不同空间模型与出租车超速行为特征的拟合情况；同时，考虑道路组间异质性，进一步构建随机系数空间模型。结果表明：出租车路段超速线密度和超速严重度均存在显著空间自相关性，空间模型 对其拟合效果普遍优于传统模型；不同超速行为特征适用不同的空间模型，随机系数空间条件自回归模型对超速线密度拟合效果最优，而不同百分位数超速严重度适用的最佳拟合模型差异较大；路段限速、一块板横断面及非机动车车道这3类因素表现出对出租车超速行为特征的组间影响异质性；因素解释方面，路段限速、一块板横断面、非机动车车道、路段施工区与超速线密度显著相关；路段限速、非机动车车道、上下坡匝道、路段施工区及路段长度与超速严重度显著相关。

关键词: 交通工程, 道路因素, 组间随机系数空间模型, 超速行为, 空间条件自回归模型, 空间同步自回归模型