Researches On Real Time Eye Tracking In Random Scenes And Application Technology

In recent years,eye detection and gaze tracking technology has become an important research topic in computer vision and pattern recognition.Because more than 50% of human information is obtained from the visual system,and human eyes are the important features of the human face.Therefore,the research of human eye location and movement is an important way to analyze human psychology and behavior.With the development of computer hardware performance,human eye detection algorithms based on computer vision have become the mainstream research method for human eyes,due to its non-invasive,easy collect images,high precision and robustness.Human eye detection and gaze tracking technology have been widely used in face detection,iris recognition,identity verification,facial expression analysis,disease diagnosis,behavior research,military investigation,auxiliary driving,artificial intelligence,virtual reality,data analysis,etc.Therefore,how to perform human eye detection algorithms efficiently and accurately are of great significance.Researchers have proposed several of human eye detection algorithms,such as electro-oculography(EOG),corneal reflection(CR),infrared sensors positioning and human eye feature extraction.However,existing human eye detection methods mostly need high resolution near-infrared or visible light camera or multiple low resolution near-infrared cameras to record eye images,The eye detection devices are complex,hard to use,limited application scenario and very expensive.The eye detection algorithms relies on the resolution of the images.In addition,the algorithms are less robust and real time for random scenes where the face is externally occluded,and the detection accuracy is low.These problems have limited the human eye detection algorithms applied in our life.Based on the existing human eye detection technologies and the characteristics of human eye structures,this paper presents a series of image-based human eye detection algorithms based on the random scenes,and establishes the low-cost human eye-gaze tracking system.The proposed human eye detection algorithms and gaze tracking systems are applied in the fields of behavior analysis and medical diagnosis.The performance of human eye-gaze tracking algorithms and systems are also analyzed and discussed.The innovations of this paper mainly include:1.In order to solve the problem of external light changes,eyelashes or eyelids occlusion,and the traditional eye detection algorithms cannot accurately locate the human eye regions and pupil centers in the infrared human eye images.The human eye-pupil location algorithm based on deformable template matching is proposed.The proposed algorithm firstly locates the human pupil regions by Haar-like features.Then,the deformable eye template is used to fit the pupil contours and center within the pupil regions.The experimental result shown that the proposed algorithm has comparable or better accuracy and is more efficient,compared with the two state-of-the-arts methods in different datasets.The pupil center location accuracy within 5 pixel error range,the detection rate can reach 84% and 91% in Swirski and our datasets.The proposed pupil center detection algorithm can be used as the feature point of the face detection,iris recognition and gaze tracking.2.In order to solve the problem that the most existing human eye detection algorithms can only locate the eye positions in the near-infrared eye images,the face images detection rate are low and the algorithms are poor in robustness in the low-resolution random scene.A hybrid human eye detection method which is based on machine learning and regression tree model is presented.The proposed algorithm can locate human eye region and pupil center in the near-infrared or visible human face images,and consider the influence of different image resolutions(minimum face image resolution up to 384?286 pixels).The proposed algorithm is tested in BioID and GI4 E datasets,and compared with several the-state-of-the art methods.The experiment results shown that the proposed algorithm can efficiently and accurately locate the human eye position and pupil center in infrared or visible low-resolution images.When the normalized error is less than 0.05,the pupil center detection accuracy both in the BioID and GI4 E datasets can reach 90.8% and 95.4%,which is superior to existing methods.3.In order to solve the problem that the human eye detection algorithms are weak in external occlusion and poor real-time in random application scenarios,a real time human eye detection method based on cascaded convolutional neural network is presented.The proposed algorithm uses a series of convolutional neural networks with similar structure to locate the left and right eye region firstly.Then,the second convolutional neural network is used to locate the eye centers.The proposed algorithm completely based on datasets training idea and generate an efficient human eye detection model by the CNNs.In order to reduce the number of parameters in the convolutional neural network and avoid the long training time due to the global search,an improved gradient pyramid model based on SIFT algorithm is proposed to simplify the number of human eye candidate regions input to the CNNs.The eye detection accuracy and efficiency of the proposed algorithm compared with existing algorithms: when the normalized error is less than 0.05,the detection rate of BioID dataset can reach 85.6%.It takes only 13 ms to process a single face image in BioID dataset.The proposed algorithm can meet the requirements of real-time human eye detection using the web-camera(frame rate is 30-60 f/s).The experimental results demonstrate that our proposed method can achieve satisfactory results both on eye region and pupil center detection on benchmark datasets4.In order to solve the problem of the current commercial gaze tracking systems are expensive and the laboratory-level gaze tracking devices have low detection accuracy,a novel low-cost gaze tracking system has been designed using the proposed human eye detection algorithms.The main innovations of the proposed system are: it is the first time uses the near-eye display technology in the gaze tracking system and integrates the proposed human eye detection algorithm.An efficient gaze tracking algorithm based on convolutional neural network combined geometric model method is proposed.The system also uses a low-cost infrared micro camera to record the human eye movement in real time and display of external scenes in the near-eye viewing device.Based on the proposed gaze tracking system,the real-time gaze tracking,eye-control keyboard and gaze positions analysis are proposed,which fully verify the proposed system’s gaze tracking accuracy can reach 0.53 degree.