EAPSI: Preventing Traffic Accidents based on Driving Behavior with Engineering and Statistical Models

预测驾驶行为是开发更安全、更可靠的驾驶辅助系统的重要组成部分。谷歌自动驾驶汽车和Mobileye系统等现代技术是使用传感器或摄像头检测道路状况并在潜在危险即将发生时向汽车和驾驶员发出指示的好例子。然而，这些技术在极端天气条件下（如大雨或夜间黑暗）行驶时可能不太可靠。因此，工程领域的许多研究人员都试图通过分析驾驶员的行为来预测接下来会发生什么。这个项目提出使用驾驶员行为数据来预测即将发生的交通事故的可能性。该项目是与日本奈良科学技术研究所的Kazushi Ikeda教授合作的。池田教授和他的实验室是分析交通数据的专家。隐马尔可夫模型（HMM）是工程领域中常用的交通数据分析技术。我们可以将HMM视为一种无监督聚类技术，它试图将不同的驾驶状态聚类成显示相似数据模式的组。利用HMM对不同的驾驶状态进行估计和预测的研究已经有很多，但是将这一思想扩展到预测事故发生的研究却很少。该项目试图通过将现有方法与时间序列分析的附加变化点检测程序相结合来缩小这一差距，从而在驾驶员行为异常时预测事故的发生。简而言之，HMM将首先用于学习不同的驾驶状态。一旦状态被估计出来，就可以得到每个状态内的一个全局趋势并去除，然后应用变化点分析来分析数据的残差，其目的是检测驱动因素的突然变化。可能预示事故即将发生的行为。这可以看作是使用无监督学习（HMM）的预处理步骤，然后使用统计建模方法（变化点检测）来构建预测模型。该奖项隶属于东亚和太平洋暑期研究所项目，由美国国家科学基金会和日本科学促进会共同资助，支持一名美国研究生进行暑期研究。

Predicting driving behavior is an important component for developing safer and more reliable driver assistance systems. Modern technologies such as the Google Self-Driving Car and the Mobileye system are good examples on using sensors or cameras to detect road conditions and send out instructions to the cars and drivers when potential dangers are about to occur. However, these technologies can be less reliable when driving in extreme weather conditions, such as heavy rain or darkness at night. Thus many researchers in the engineering field have attempted to analyze driver behavior to predict what will happen next. This project proposes the use of driver behavior data to predict the possibility of an upcoming traffic accident. The project is in collaboration with Professor Kazushi Ikeda from Nara Institute of Science and Technology in Japan. Professor Ikeda and his laboratory are experts in analyzing traffic data.Hidden Markov Models (HMM) is a popular technique used in the engineering field to analyze traffic data. One can view HMM as an unsupervised clustering technique, where one tries to cluster the different driving states into groups exhibiting similar data pattern. Much work has been proposed on estimating and predicting the different driving states using HMM, however, little work has extended this idea to predicting the occurrence of accidents. This project attempts to close this gap by combining existing methods with an additional change point detection procedure from time series analysis to predict the occurrence of an accident if the driver behaves abnormally. In brief, HMM will first be used to learn the different driving states. Once the states are estimated, a global trend within each state can be obtained and removed, then change point analysis will be applied to analyze the residual of the data, where the goal is to detect sudden changes in drivers? behavior which might signify an accident is about to occur. This can be viewed as a preprocessing step using unsupervised learning (HMM) followed by a statistical modeling approach (change point detection) to build the prediction model.This award under the East Asia and Pacific Summer Institutes program supports summer research by a U.S. graduate student and is jointly funded by NSF and the Japan Society for the Promotion of Science.