集成多模态信息的驾驶者异常状态识别模型研究

项目名称： 集成多模态信息的驾驶者异常状态识别模型研究

项目编号： No.51308096

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 建筑科学

项目作者： 杜勇

作者单位： 东北农业大学

项目金额： 25万元

中文摘要： 目前基于视觉信息的驾驶者异常状态识别方法主要依靠彩色图像进行分析，现有的研究大多只关注单一区域特征，这样限制了能够识别异常状态的种类，加之对序列信息关注不足，以及客观环境如车内背景、颠簸以及光线干扰，使得分类的可靠性证难以保证。这是目前基于视觉信息进行驾驶者状态识别中存在的两大问题。 本研究旨在解决上述难点问题，本课题首先对监控信源进行改进，结合多红外视频和音频采集装置共同监控驾驶者的状态，从而保证获取信息的可靠性和互补性。课题主要从以下三方面展开研究，首先，在各关键区域针对不同信源提取与驾驶者状态相关的多模态描述特征构建原始特征集合。其次，借助于粗糙集技术对原始特征集合进行评价进而实现对原始特征集合的约简。最后，在约简后的多模态特征集合上进行集成核分类器学习，并利用习得的集成核分类器对驾驶者状态进行识别。

中文关键词： 模式识别；驾驶异常状态识别；特征评价；稀疏表示；深度学习

英文摘要： At present, most research on driver abnormal state detection depends on color images. Most of the existing research only concerns with the features of a single region. This restrict the detection of different kinds of abnormal activities. Furthermore, coupled with lack of concern of sequence information, as well as some objective factors such as background, jolt in the car and light interference, it seems difficult to ensure the reliability of classification. These are two main issues existing in the visual information based driver state recognition task currently. This research aims to address the above problems. First of all, the monitor is improved. Infrared cameras and sound recorders are used to monitor the state of the driver, so as to ensure the acquired information to be reliable and complementary. The research is carried out from the following three main aspects. Firstly, extract multi-modal information to describe driver states from each critical driver-status ralated areas and build original feature set. Secondly, through feature evaluation, the original feature set is reducted based on rough set techniques. Finally, a ensemble kernel classifier is learned on the selected multi-modal feature subset and the achieved ensemble kernel classifier is used for the driver state recognition.

英文关键词： pattern recognition；driving abnormal state；feature evaluation；sparse representation；dimension reduction