Research On Driver Fatigue Status Classification Via Human Facial Information

With the development of society and economic, automobile has become an essen-tial transport tool for expanding living space, improving the efficiency of life as well asimproving life quality. The soaring number of vehicles reflects the prosperity of the so-ciety, at the same time it brings a lot of social problems.First is the environmental andtraffic congestion problems in cities, however, the primary issue is the current road safetyproblems. According to statistics, road traffic accidents worldwide each year is up to10billion, accounting for90%of the total number of security incidents worldwide or so, thenumber of injuries is about25million and caused at least500,000deaths as a result ofroad traffic accidents, accounting for more than80%of the global safety accident casu-alties. Road traffic accidents caused by human beings has become a primary factor inunnatural deaths. Each country has to pay a heavy price on human lives and enormouseconomic losses in road traffic accidents per annum.However, unfortunately traffic environment in our country are even worse, the trafficaccident rate has ranked first in the world for years. In China, the annual number of deathsin road traffic accidents accountes for more than20%of the world’s. For such a badtraffic environment, improving road safety seems imperative. Statistics show that in allroad traffic accidents, human factors account for80%，while fatigue driving is the mostcommon human factor. Therefore, the identification and early warning for fatigue drivingwill play a crucial role in avoiding the occurrence of malignant accidents, guaranteeingpeople’s lives and property. The research content of this paper just carry out around theissue of identification of fatigue driving.First, the significance of the study on the subject of identification of fatigue driv-ing is well explained and the current research status at home and abroad is summarized.Next, different identification techniques of fatigue driving is introduced in detail, at thesame time, various identification techniques are classified and summarized. Four types ofcommonly used objective detection means are focused on and the advantages and disad-vantages of them are compared. Finally, we further conclude the future trends in the fieldof identification of fatigue driving as well as summarize the research hotspots at homeand abroad through the analysis of important documents related. Meanwhile, in this por- tion of the thesis, it is declared that the research is based on the visual information of thedriver.Second, the issue of location of human face and its local units in fatigue drivingidentification is well discussed. By summarizing and comparing the related methods,eventually two face and its local units detection methods adapting to the driving envi-ronment come into being. One is skin color Gaussian model combining with templatematching for face region and local organs detection. The other is Haar-like features basedAdaBoost framework combining with intensive facial image for face region and local or-gans detection. Through eyes and mouth position, a number of fatigue-related facial unitscan be easily located.Third, our research begin with two kinds of significant fatigue performance and twofatigue detection methods are proposed accordingly. The first method is to detect theremarkable performance that when fatigue occurs the blinking process becomes signifi-cantly slower. The difference between awake blinking and fatigue blinking processes inthe frequency domain is analyzed in order to find a specific frequency band which candistinguish two blinking processes efficiently. Then the fatigue index during the timeis computed on this specific frequency band to characterize the driver’s fatigue degree.The significance of this approach is providing an new important perspective to describethe fatigue of the eyes. To a certain extent it can compensate the Perclos method lacks.The second method focuses on another remarkable performance of fatigue, namely yawn.Three states of the mouth, namely silence, talking and yawning, is defined aiming at thelack of past researches. At the same time, we point out that to be effective in yawningclassification yawn must be classified as a process. So sequence features are introducedin classification of yawning and the yawning classification problem is well solved.Fourth, a multi-observation areas based ensemble learning model is proposed forgeneral fatigue performance identification. It is believed that although the expression offatigue is not obvious, if the observation areas are concerned, the performance of fatiguewill be relatively concentrated and the law of expression changes will be much clearer.Another advantage of feature extraction from the observation areas is that it can greatlyreduce the interference caused by redundant information in face when learning classifiers.For each observation area a C4.5base classification model is built. Each base model isequivalent to an independent decision maker. But due to its limited capacity, it may not be able to give accurate decisions. However, if the information provided by each decisionmaker is combined together, it will form a comprehensive evidence. The driver statusclassification results given by ensemble learning approach are more accurate and stable.Finally, the nature of C4.5decision trees good performance in the classification onfatigue data sets is analyzed and it is concluded that rule learning classifier is more suit-able for the classification of sparse distribution samples in high-dimensional nonlinearfeature space. So a novel rule learning classifier with more better generalization abilityis proposed. It is named neighborhood covering reduction classifier. This classifier inte-grates the inherent powerful approximation ability of rough set in approaching arbitraryconcepts. Supported by reasonable covering elements reduction and pruning strategy, asimple and compact rule set can be obtained for classification and it has a relatively bettergeneralization ability than C4.5decision tree. The classifer is applied in our classificationtask and find that the proposed classifier has even better classification results than someother of the classical classifiers, including NEC, NN, LVQ, LSVM, CART, and Jripper.Besides constructing the neighborhood covering reduction classifier a set of theoreticalframework of relative covering reduction is established correspondingly which has im-proved and supplemented the theory of covering rough set.