已有的交通事故黑点鉴别研究大多基于事故频数或事故率，并未考虑交通事故对不同发生地的影响特征。为了综合考虑交通事故在不同交通环境和路网特征下的差异影响，并解决交通事故数据中零值远超经典离散分布的零膨胀问题，本文提出一种考虑节点综合重要度的改进网络核密度估计法，并基于零膨胀负二项回归模型对城市交通事故黑点进行鉴别。首先，在拓扑路网中综合考虑事故发生地的交通环境和道路条件构建事故综合影响度指标，连同事故严重程度指数嵌入到传统网络核密度估计中，通过在道路网络上生成平滑的密度表面定性体现点事件的空间聚集性。在此基础上，构建基于零膨胀负二项回归鉴别模型，明晰事故高发区域边界范围，定量刻画不同严重等级的事故黑点路段空间分布特征。最后，以深圳市华强北街道为例进行实例分析。结果表明，在90%、80%和70%的阈值水平下本文提出的事故黑点鉴别法的有效搜索率均高于平面核密度估计法，且考虑节点综合影响度后，部分无道路区域不再被误识，模型准确率比传统网络核密度法分别提升了3.60%、5.31%和7.20%。

The existing research on identifying black spots in traffic accidents is mostly based on accident frequency or accident rate, without considering the impact characteristics of traffic accidents on different locations. In order to comprehensively consider the differential effects of traffic accidents in different traffic environments and road network characteristics and to solve the zero inflation problem of zero values far exceeding the classical discrete distribution in traffic accident data, this paper proposed an improved network kernel density estimation method that considers the comprehensive importance of nodes, and identified urban traffic accident black spots based on the zero inflation negative binomial regression model. Firstly, in the topological road network, a comprehensive impact index of accidents was constructed by comprehensively considering the traffic environment and road conditions at the location of the accident, and the accident severity index was embedded into the traditional network kernel density estimation. By generating a smooth density surface on the road network, the spatial aggregation of point events was qualitatively reflected. On this basis, a discrimination model based on zero-inflated negative binomial regression was constructed to clarify the boundary range of accident-prone areas and quantitatively depict the spatial distribution characteristics of accident black spots at different severity levels. Finally, an example analysis was carried out for Huaqiangbei street in Shenzhen. The results show that the search efficiency indexes of the proposed method are all larger than those of the planar kernel density estimation method at the threshold levels of 70%, 80% and 90%. Furthermore, some non-road areas are no longer mistaken, and the accuracy of the model is 3.60%, 5.31% and 7.20% larger than those of the traditional network kernel density method respectively after considering the comprehensive importance of nodes.