Research On Eye Movement Recognition Based On Convolutional Neural Network

In the field of fatigue detection and human-computer interaction,eye movement recognition is one of the key technologies and has attracted more and more attention in the past ten years.However,traditional eye movement recognition algorithms often require additional devices to achieve contact data collection.In addition,theses algorithms are sensitive to illumination and obstacles which lead to the degradation of system performance.Convolutional neural network is a deep learning-based pattern recognition method which develops in recent years.Convolutional neural network is widely employed in image recognition due to its excellent location perception and high-level polymerization.Compared with traditional image recognition algorithms,convolutional neural network can extract features from image automatically,without depending on specific manual features.In addition,convolutional neural network has better recognition performance and generalization ability.Therefore,in this thesis,we proposed an eye movement recognition method based on convolutional neural network and investigated its recognition performance and robustness.The algorithm was employed in fatigue detection.Moreover,in order to establish an effective eye movement image dataset,we also investigated the human eye detection algorithm.The main work is as follows:(1)Research on human eye detection algorithm based on AdaBoost.According to the first rectangular feature of AdaBoost-based face classifier is eye area,AdaBoost-based eye classifier was employed in this rectangular feature to detect human eye,which could narrow the search area of human eyes and improve the efficiency of the algorithm.The AdaBoost algorithm mainly learns about the features of samples which greatly speeds up face detection and makes it possible to apply face detection in practical applicatiopns.(2)The eye movement datasets were built by selecting 24000 eye movement images manually.The dataset were labeled in accordance with the requirements of the convolutional neural network model.(3)Research on eye movement recognition algorithm based on convolutional neural network,and 2-class,4-class and 8-class eye movement models were tested respectively.Compared with other eye movement recognition algorithms based on BP,SVM,and EOG,the experimental results demonstrate that convolutional neural network is a high-performance eye movement recognition algorithm with strong noise resistance.The recognition results show that the recognition rate of the proposed eye movement recognition algorithm is higher than other common eye movement recognition algorithms.The algorithm can still maintain good recognition performance under the environment of strong noise interference such as illumination.Moreover,the influences of structural parameters of convolutional neural network were studied,such as the activation function,the size and number of convolution kernels and the number of convolutional layer.In this thesis,the eye movement recognition algorithm based on convolutional neural network was employed in the doze detection.We selected optimal network structure parameters to construct eye movement recognition model.The changes of opening and closing of the eyes were detected.According to the detection results,the subject's fatigue was estimated based on PERCLOS principle.(4)The construction of convolutional neural network-based eye movement recognition system.According to the convolutional neural network model that trained by eye movement datasets,we established an eye movement recognition system on Python by using PyQt and tensorflow.