Research On Human Activity Recognition Algorithm Based On EOG In Office Scene

Human Activity Recognition(HAR)can be defined as measuring and identifying comprehensive information such as the types of reflection and patterns of behavior,and the recognition results can be presented by a naturallanguage.Because of the ability of user’s intention perception,the HAR system has wide application prospect and potential commercial value in the fields of advanced human-computer interaction,intelligent video surveillance,virtual reality and medical diagnosis.Nowadays,most of the researchers focused on daily activities(such as running,walking,jumping,upstairs,downstairs,etc.),which can be expressed by the body’s movement.However,there are few reports about visual activities recognition relating to eye movement without obvious limbs movement(such as reading,writing,watching video and browsing the web).As an important recording method of eye movements,EOG signal has the advantages of non-invasive,low cost,easy to carry and less influence of illumination.Therefore,this thesis focuses on activities recognition in the office scenes using EOG signals.The specific works include:(1)Investigate the existing human activity recognition technologies,in addition to analyze and compare implementation method based on computer vision and bioelectric sensors.On this basis,according to the characteristics which recognition activities(reading,resting,and writing)are mainly related to eye movement,the EOG signal is selected as the method of recognition.Finally,the general EOG-based signal processing methods including acquisition,preprocessing and feature extraction are studied.(2)Research on determination and recognition of basic eye movement units.During the procedure of EOG-based activity recognition,the duration of eye movement unit is important for the recognition of the system.To explore the relationship between the duration and recognition ratio,we first collect eye movement data with 6 different time periods(i.e.,5s,10s,15s,20s,25s,and 30s).Then,we use three traditional methods(i.e.,wavelet transform,power spectral density,and Hjorth parameters)to extract feature parameters for identifying the aforementioned 6 basic eye movement units.Finally,we determine the duration of 10s as the optimum basic eye movement unit by comparing the recognition ratios under different time lengths.The conclusion lays the foundation for the following studies.(3)Research on the establishment of activities relationship model.Under a specific background task,there exists a potential contextual relationship between different behavior states.Reasonable usage of this relationship may improve the performance of the human activity recognition.Therefore,in order to establish an effective activities relationship model,we first design a new experimental paradigm to collect eye movement data under different tasks.Subsequently,we apply the N-gram algorithm to calculate the transition probability of each behavior state to infer the current state of behavior in terms of the previous state sequence.(4)Research on human activity recognition with dual model fusion.In order to validate the effectiveness of the proposed activities relationship model,we first propose a confidence parameter,which used to integrate the outputs from activity relationship model and EOG signals recognition model.This confidence parameter to judge whether the recognition result of EOG signals recognition model is high risk and then re-identify these high-risk results by using the probability statistical information obtained from the activity relationship model to improve the reliability of the result.Under the within-subjects test(training data and the testing data come from a same subject)and the between-subjects test(training data and the testing date come from different subjects),the average recognition accuracy ratios of combining the activities relationship model obtains an obvious improvement of 7.72%,2.49%,4.69%and 11.76%,12.1%,12.42%compared with traditional EOG signal recognition model.The experimental results reveal that the activities relationship model can indeed improve the performance of human activity recognition algorithm.