Study Of Driver’s Eye Detection And Eye State Recognition Algorithm Under Complex Environment

Traffic accidents caused by driver fatigue accounted for a proportion which cannot beignored. Lots of researches on how to detect driver fatigue were launched by nationalresearchers. Currently, an eye feature-based detection method, is considered the most effective,time-saving and relevant driver fatigue detection method. Eye location and eye statesrecognition under complex environments are two key technologies in driver fatigue detection.Image acquisition system which includes850nm LED light source and a narrowbandfilter was used in this paper to reduce the impact of complex illumination and meet therequirements for night use. On this condition, this paper studies the eye localization undercomplex environment, set up a complete eye-image databse under near infrared condition,extracts features which can effectively describe the state of eye, constructs an eye staterecognition model under complex environment.The main work of this paper is as follows:(1) In the driver fatigue detection system, both open eyes and closed eyes should belocated. Changing illumination, glasses shelter, imaging conditions make applicationscenarios more complicate, so using a single feature or location method cannot guarantee therobustness of location. After comparison of three location methods respectively based on Haarfeature, blob detection and active shape model, a location method based on a cascade of thelocation methods based on Haar feature and blob detection was proposed. In the near-infraredimaging conditions, this method can locate both open eyes and closed eyes. Without glassesand open eye with glasses, this method achieves high location accuracy. With sunglasses andclosed eye with glasses, this method can also locate eyes mostly. What’s more, the lowcomputational complexity of this method meets real-time requirements.(2) After the precise location of eyes, the eye states recognition was required. With thechange of facial expression, head posture and the block of ornaments, the shape of eye wouldchange correspondingly which brought a great challenge. This paper constructed a completenear infrared eyes image database used for training and testing. The database contains38kinds of different patterns of changes, a total of more than hundreds of sample images. Thedatabase contains rich change mode, so it is convenient for researchers to study the key problem of algorithm.(3) A near-infrared eye states recognition algorithm based on HOG-LBP feature fusionwas proposed. The HOG feature can represent the gradient of eyelid and pupil region well,and the LBP feature can represent the local texture of eye well. After the PCA (PrincipalComponent Analysis) of these two features to reduce the dimensions, they were fused intoHOG-LBP feature. Lastly, eye states recognition model was built which adopted SVM(Support Vector Machine) based on RBF (Radial Basis Function). The proposed algorithmperforms well in both19modes of open and closed eyes. The average accuracy rate of openeyes is95.73%and the average accuracy rate of closed eyes is97.16%. The averageprocessing time of this method can also fully meet real-time requirements at low processingplatform.