Millions of motor vehicle collisions occur each year and lots of them result in heavy fatalities. Although some promising works are proposed, they have the following problems: (1) most of existing methods depend on feature regression, but ignore the causal relationship among them; (2) the vision-based techniques cost enormous resources to process the large scale of video data; (3) the lack of considering real-time traffic environment leads to an unsatisfied performance. To tackle these problems, we propose a real-time explainable collision inference framework through social media analysis. First, we design and extract various kinds of real-time traffic features from the social media. Then, we propose an effective algorithm to discover the causal relationships among the adopted features, which are denoted by probabilistic graphs. Finally, we employ the probabilistic graphs with the top-$k$ BDeu score to calculate the probability of one collision occurring with nearly linear time complexity. Extensive experiments show that our framework achieves 0.752, 0.747, and 0.750 in precision, recall, and F1-measure. Extensive results show that our proposal has good scalability and has a good chance to solve other emergency event inference.

每年发生数以百万计的汽车碰撞，其中许多导致严重的死亡。尽管提出了一些有希望的工作，但它们存在以下问题：（1）大多数现有方法都依赖特征回归，而忽略了它们之间的因果关系；（2）基于视觉的技术需要花费大量资源来处理大量视频数据；（3）缺乏考虑实时交通环境导致性能不满意。为了解决这些问题，我们通过社交媒体分析提出了一种实时可解释的碰撞推理框架。首先，我们从社交媒体设计和提取各种实时交通特征。然后，我们提出了一种有效的算法来发现采用特征之间的因果关系，这些概率关系用概率图表示。最后，$ķ$BDeu分数用于计算发生一次碰撞的概率，时间复杂度接近线性。大量的实验表明，我们的框架在精度，召回率和F1度量上达到了0.752、0.747和0.750。大量结果表明，我们的建议具有良好的可扩展性，并且有很好的机会解决其他紧急事件推断。