Eye Detection Based On Local Features

Eye detection is a necessary step for iris recognition,face recognition,eye state estimation,fatigue driving detection,eye tracking,etc.In recent years,eye detection has become a research hotspot in computer vision and pattern recognition,and it has significant value of academic research.With the rapid development of artificial intelligence,application prospects of eye detection become more and more extensive,including psychology,video surveillance,virtual reality,augmented reality,human-computer interaction,etc.Eye detection is often affected by factors such as illumination,accessories,occlusion,complex background,and so on.Thus,it is extremely important to study features which are extracted to reduce the impact of the above unfavorable factors as much as possible and achieve real-time and efficient detection speed.This thesis focuses on the problem about real-time and effective eye detection under the influence of factors such as illumination,wearing glasses,closing eyes,glare of glasses and so on.The major distributions of the thesis are included as follows:(1)In order to locate eyes rapidly and accurately,a new local feature descriptor,multi-scale local block LBP histogram(MSLB-LBPH),is present.It is fast to calculate MSLB-LBPH.Furthermore,MSLB-LBPH has a strong ability to describe local features of eye images and is especially suitable for training a cascade classifier based on Real Adaboost algorithm;(2)An eye detection algorithm based on MSLB-LBPH is proposed.Firstly,eye images as training samples are normalized and preprocessed.And MSLB-LBPH features are extracted.Then Real AdaBoost algorithm is used to train a cascade classifier,which can detect eyes in normalized facial images effectively;(3)Expanding on an eye detection algorithm based on MSLB-LBPH,another eye detection algorithm based on MSLB-LBPH and Co-HOG is proposed.Co-HOG expresses the information of eye image shape,and has strong robustness under illumination and deformation.Extensive experiments have been carried out on the public face databases.The experimental results show that proposed method is superior to the latest other methods and has some advantages: high accuracy,fast detection speed,and robustness.