基於腦波之駕駛員認知反應估測及其在安全駕駛的應用

Abstract

在這篇論文裡，我們發展先進的生醫訊號處理技術，結合獨立成份分析演算法、腦波頻譜分析、相關係數分析及模糊類神經網路，在虛擬實境技術所建構的認知駕駛實驗環境中，透過非侵入式記錄的腦波電位分析，來辨識人腦對事件刺激的暫態反應以及估測駕駛者的精神警覺程度。並且應用此分析技術去動態量測駕駛員的精神認知狀態變化，與相對應的認知、辨識、車輛控制能力及駕駛行為的改變，以維持駕駛員的高效能表現，並防止駕駛員在開車時，因過失及失誤所造成的意外事故。 我們首先提出一個新的基於獨立成份分析之時態匹配濾波器，去分析單次事件相關腦電位，過濾人為雜訊的干擾，並改進傳統時域疊加方法的缺點，最後利用模糊類神經網路模型去識別駕駛員看到紅綠燈號誌所產生的相對應腦波暫態反應。實驗結果驗證利用多維腦電位訊號去辨識人的精神認知狀態與對事件刺激所造成的腦波暫態反應是可行的，我們所提出的方法可以提高所量測之腦波暫態反應的訊號品質，以達到較高的辨識率。 我們也提出一個新的基於獨立成份分析之適應性特徵值選擇機制，可以從腦波頻譜中選取最有效的獨立成份及具代表性的頻段作為特徵值，對每一個駕駛員建立個別的模糊類神經網路模型，去探索在疲勞或失神時，腦波的活動特性以及所伴隨的駕駛行為變化。實驗結果顯示，人的腦波頻譜變化與開車行為表現之間的關係非常密切，我們所提出的方法不但可以去除大部分人為雜訊的干擾，並且能夠估測最佳的頭皮位置來放置腦波偵測器，以精確的估測駕駛者的精神警覺程度與實際的開車行為變化。此腦波訊號分析技術未來可以利用可攜式嵌入式系統來實現一個線上精神狀態監控系統，以應用到日常生活中。

In this thesis, we develop advanced biomedical signal-processing technologies that combine independent component analysis (ICA), power spectrum analysis, correlation analysis, and fuzzy neural network (FNN) models to assess the event-related transient brain dynamics and the level of alertness of drivers by investigating the neurobiological mechanisms underlying non-invasively recorded electroencephalographic (EEG) signals in the virtual-reality-based cognitive driving tasks. The developed techniques are then applied for dynamically quantifying driver’s cognitive responses related to perception, recognition, and vehicle control abilities with concurrent changes in the driving performance to maintain their maximum performance in order to prevent accidents caused by errors and failures for driving safety. We first propose a novel ICA-based temporal matching filter for analyzing the single-trial event-related brain potentials (ERP) without using conventional trial-averaging results as input features of the FNN classifiers and apply this method to recognize the different transient brain responses stimulated by red/green/amber traffic-light events. Experimental results demonstrate the feasibility for identifying multiple streams of EEG signals related to human cognitive states and responses to task events. Our proposed methods can dramatically increase the quantity and quality of momentary cognitive information and achieve high recognition rates. We also develop a new ICA-based adaptive feature-selecting mechanism to extract most effective bandpower from EEG power spectrum and build an individual FNN model for each subject to further examine the neural activities correlated with fluctuations in human alertness level accompanying changes in the driving performance in the lane-keeping driving tasks. Experimental results show a closed relationship between changes in EEG power spectrum and the subject’s driving performance. Our proposed models also can effectively remove most non-brain artifacts and locate optimal positions to wire EEG electrodes such that it is possible to accurately estimate/predict the continuous fluctuations in human alertness level indexed by measuring the driving performance quantitatively. The computational methods are well within the capabilities of modern digital signal processing hardware to perform in real time and thus might be used to construct and test on a portable embedded system for an online alertness monitoring system in the future.