Abstract: Since fatigue driving is one of the main inducements of traffic accidents, it is of great application value to explore objective and accurate method detection for fatigue driving. Considering the information complementary between different types of features, as well as the advantages complementary between different machine learning algorithms in the process of information mining, this paper proposes a fatigue-driving detection framework based on spatial-temporal electroencephalogram (EEG) features and parallel neural networks. To reduce the volume conductor effect, map the time series EEG data to the spatial brain functional network (BFN) based on the phase-locked value (PLV), and successively extract temporal domain EEG features and spatial metric features related to the driving process from the time-series EEG data and BFN. A feature contribution algorithm is designed by analyzing the relationship between features and target classes, to give different contribution coefficients for the temporal domain EEG features and the spatial domain BFN metric features. And the two types of weighted features are used as the inputs of the long short term memory (LSTM) network and the two-dimensional convolutional neural network (2D-CNN), to use the advantages of LSTM network in temporal data processing and CNN in the spatial data processing, and thus realizing the complementary information of spatial-temporal EEG features and the complementation of two types of neural network algorithms in data mining ability. A series of comparative experiments on public datasets show that the fatigue detection performance of the parallel neural network framework is superior to other methods, and the highest detection accuracy is 96.47%. This result means that this method can provide an effective solution for fatigue driving warning and assist safe driving.

Key words: intelligent transportation, fatigue driving detection, spatial-temporal electroencephalogram features, safe driving, neural network, brain functional network

摘要： 鉴于疲劳驾驶是交通事故的主要诱因之一，探索客观准确的疲劳驾驶检测方法具有重要应用价值。考虑到不同类型特征之间的信息互补，不同机器学习算法之间在信息挖掘过程中的优势互补，本文提出一种基于时空脑电(Electroencephalogram, EEG)特征与并行神经网络的疲劳驾驶检测框架。减少容积导体效应，基于锁相值(Phase Locked Value, PLV)将时序EEG数据映射到空间脑功能网络(Brain Functional Network, BFN)，先后提取时序EEG数据和BFN中与驾驶过程相关的时域EEG特征和空域度量特征。通过对特征与目标类之间关系的分析，设计特征贡献度算法，为时域 EEG 特征和空域 BFN度量特征赋予不同的贡献系数，分别将两类加权特征作为长短期记忆(Long Short Term Memory, LSTM)网络和二维卷积神经网络(Two-dimensional Convolutional Neural Network, 2D-CNN)的输入，充分发挥LSTM网络时序数据处理优势和CNN空间数据处理优势，实现时空EEG特征信息互补以及两类神经网络算法数据挖掘能力的优势互补。在公开数据集上进行系列对比实验，结果表明并行神经网络框架的疲劳检测性能优于其他方法，获得了最高96.47%的准确率。此结果意味着本方法能够为疲劳驾驶预警和辅助安全驾驶提供一种有效的解决方案。

关键词: 智能交通, 疲劳驾驶检测, 时空脑电特征, 安全驾驶, 神经网络, 脑功能网络