Intention Recognition Via Eye Movement Information

As the population ages,there is an increasing need for care for the elderly and disabled.At present,the professional nursing staff is in great shortage.Families and society are under tremendous pressure to provide care.In recent years,service robots are gradually being used in family life to help the elderly to complete daily tasks and improve their quality of life.In order to achieve natural and effective human-robot interaction,robots need to actively recognize human intentions.Existing methods of intention recognition usually use speech,gesture,or physical contact for human-robot interaction.However,due to the reduced abilities of speech,behavior or motor expression of the elderly and disabled,human-robot interaction is inefficient and intention recognition is difficult.Eye movement is an efficient way for human-robot interaction which can reduce the physical and psychological burden of the elderly.Therefore,this thesis proposes an intention recognition framework based on eye movement information to recognize the intention of the elderly.By analyzing eye movement features,people’s intentional eye movement is detected,and the objects of intentional eye movement are obtained.The recognition of intention is achieved using probabilistic models.The main research contents are as follows:Firstly,an eye movement dataset(Eye Movement Features,EMF)is established in a family kitchen scene.The eye-tracking device Tobii Eye Tracker 5 was used to obtain the time and position information of the user when they are looking at the interactive interface.Four eye movement features,namely fixation time,fixation count,gaze interval and gaze speed,are selected to classify the eye movement into intentional eye movement and unintentional eye movement using the SVM classifier.Based on the intentional eye movements,we obtain information about the objects that the user intends to look at.The effectiveness of the intentional eye movement detection model is verified by comparing the recognition performance on the eye movement dataset EMF.Secondly,the Naive Bayes model is used to model the probabilistic relationship between objects and intentions to recognize human intention.An object-intention dataset OIntention4 containing 14 commonly used kitchen objects and 4 daily intentions is established.Based on the object-intention dataset,the probabilistic relationships between 14 objects and 4 intentions are calculated using the Naive Bayes model.On this basis,based on the independence between each object,intentions are recognized by calculating the probability value of each intention represented by a sequence of objects.The validity of the Naive Bayes-based intention recognition model is verified by experiments.Finally,an intention recognition model based on the Hidden Markov Model is proposed.According to the different functions of object types,eight kinds of intentions in the kitchen scene are recognized.Due to the random and unobservable characteristics of human intention,this thesis uses the Hidden Markov Model to model the probabilistic relationship between the object sequence and intention sequence.An object-intention sequence dataset S_OIntention8 containing 14 kitchen objects and 8 kinds of intentions is established.Based on this dataset,this thesis takes the object sequence as the observable sequence,and the intention to be recognized as the state sequence.The hidden Markov model parameters are calculated based on the maximum likelihood estimation method,and the intention recognition is realized based on the Viterbi algorithm.In this thesis,experiments are conducted on the object-intention sequence dataset S_OIntention8 to verify the validity of the intention recognition model based on the Hidden Markov Model.