Research Of Eye States Recognition Method Based On Computer Vision

The eye output information is an extremely stable human body characteristic information.Many scholars have devoted a lot of deep research to the field of eye detection and state recognition.In daily life,military research,security and other scenarios,the combination of eye state detection and pattern recognition provides important feature information for many realistic applications,so that the robust eye state recognition shows great value in application engineering.In the past studies,eye state recognition methods are mostly processed in combination with high-resolution eye images acquired by professional equipment.However,in the trend of mobile scenarios becoming mainstream,the research results could not be well applied.Therefore,the academic research on the ‘unconstrained’ eye state recognition method is worth of more discussion.Based on image processing and classification methods with machine learning,this thesis studies a series of problems about eye state recognition in natural light images.The details are as follows.1.The thesis studies eye detection based on the improved Adaboost algorithm.When training the classifier using Adaboost algorithm,the positive and negative samples at the critical point have the weight imbalance caused by multiple misclassifications in the case of increasing number of iterations.This thesis suggests an improved Adaboost algorithm,and designed a multi-level classifier model with threshold segmentation.It effectively solves the problem of missed detection and false detection of samples in eye detection and has better robustness.2.Pupil center localization is also on the studying list.For the eye diagram,the method of center localization combined with coarse localization in eye area was used to fastly obtain the pupil center coordinates.It first roughly locates the eye area by integral projection.Then,the thesis brings up the improved Hough transform to quickly locate the pupil center on the basis of gray value and circle property on the previous result.Compared with the traditional Hough algorithm,the improved algorithm improves the computing speed of the system by about 40%,and the real-time response of the application scenario requirements is realized.3.The thesis defines eight eye-movement states,and uses the obtained pupil center point coordinate sequence as input data to realize eye movement state recognition.For the classification of multi-state problems,the state training classification was carried out by using the support vector machine based on error correction coding,and the fault tolerance of error correction coding was combined to achieve a good recognition effect.The state recognition rate of the defined eight categories can reach more than 90%.What’s more,the open,close eyes and blink behaviors of the eyes were judged.In order to verify the performance of the algorithm in this thesis,the experimental method and result analysis are given in detail.Experimental results show that the method proposed in this thesis can realize accurate eye state recognition,and the detection and recognition results can be used as output signals in various scenarios.