Research On Recognition Method Of Human Periodic Activity Driven By Single Sensor Data

In recent years,human activity recognition has become a hot area in the field of machine learning research.To better serve humans in different application scenarios by analyzing user activity information through computers.Among them,computer vision research occupies the leading position.However,in recent years,with the development of various types of wearable sensor technologies,the recognition of behaviors by using inertial sensors(e.g.,accelerometers and gyroscopes)has gradually become a hot research project in this field.Recognition of human daily behavior includes the collection and processing of raw data,the selection and extraction of features,and the design and decision-making of classifiers.Among them,the feature engineering is one of the core issues.Extracting effective features can improve the recognition accuracy,simplify the algorithm and improve the stability of the algorithm.In the traditional research of human activity recognition based on inertial sensors,the features often use the statistical information of acceleration and angular velocity,such as the mean value,variance,amplitude,kurtosis,skewness,correlation coefficient and other discrete features.The daily human movement is continuous and cyclical,discrete data does not reflect the continuity of human movement.Therefore,this dissertation focuses on the cyclical behavior of the human body,and applies it to the study of human cycle behavior recognition through the study of continuous activity by combining the reasonable features and the appropriate classification algorithms.On the one hand,the movement is continuous during the routine activity while the data collected by the motion capture system is discrete.By analyzing the continuity and periodicity of daily human activity,we propose a single sensor daily behavior recognition method based on the functional data time series modeling.Firstly,we use the functional data analysis method to fit the motion capture data of the cyclical daily activity and extract the fitted single period data.Based on this period data,we establish a hidden markov model that describing each of the daily activity processes.Finally,the maximum likelihood estimation is used to determine the discrimination results.By our method,eight kinds of routine activity can be quickly and effectively identified just by a single waist sensor.On the other hand,in the research of human activity recognition based on inertial sensors,feature extraction is one of the key aspects.The stability of discrete data statistical features depends on the window size of feature extraction.In general,the window length of the training data needs longer than one motion cycle.Therefore,aiming at the problem of short sequence sample identification whose test data is far less than one motion cycle,we proposes a new solution based on template matching.Firstly,construct an over-complete short-time behavior template library by properly segmenting the long sequence samples of the training data.The short-time samples that will be tested and the standard samples in the template library are uniformly processed and matched,Then,we obtain the classification and recognition results.Secondly,in the matching algorithm,the similarity histogram was obtained by utilizing the sum of the F-norm distance between the samples and the 2-norm distance of the global gradient vector as the matching metric.Finally,based on the similarity histogram,the final classification recognition result was obtained according to the voting strategy.Experiments show that in the case of using a single sensor to identify short-time behavior,the new algorithm had better performance than traditional algorithms in accuracy and stability.