Implicit Intention Inference Based On Gaze Series

With the advent of the AI era,auxiliary robots have entered the lives of people with limited mobility.The human visual gaze contains key and potential implicit intention information.Accurate recognition and inference of the intention expressed by visual information is an important way for intelligent robots to effectively serve humans.Based on this,this thesis starts from the visual attention mechanism,uses the Markov regime switching model model to identify gaze information features,selects the K-means clustering algorithm to extract gaze center sequences,and combines the Coherent point drift algorithm to identify gaze object sequences.The intention database is constructed through experiments,and human implicit intentions are inferred based on the intention database and Hidden Markov models.The main research content and innovation points of this thesis are as follows:1)Identifying the attention sequence based on the Markov regime switching model.In order to identify the attention state in eye movement data,the Markov region transition model is used to characterize the process of eye movement changes,and the attention points in the attention state are extracted based on the probability values in the filtering vector.The experimental results show that the average accuracy of the Markov regime switching model in identifying the attention point sequence is 91.5%.2)Recognition of gazing object sequences based on point set registration algorithms.In order to identify the object sequence of human attention,the Coherent point drift algorithm is used to register the attention center sequence extracted by the K-means clustering algorithm with the object center sequence.The experimental results show that the Coherent point drift algorithm achieves accurate matching between the object sequence of attention and the trajectory of attention information.3)Inferring implicit intentions based on multiple parallel Hidden Markov models.In order to describe the parameter features of the intent library as detailed as possible,multiple parallel single-intent Hidden Markov models are trained using the Baum-Welch algorithm.Then,the matching score between the object sequence of attention and the single-intent Hidden Markov model is calculated based on the Forward-backward algorithm,and the implicit intention of human visual attention object sequence is inferred through the matching score.The experimental results show that multiple parallel single-intention Hidden Markov models have completed the inference of human hidden intention with an average accuracy of 98.49%.