Research On Machine Learning Algorithm For Human Behavior Recognition

At present,people use embedded sensors to accumulate a large number of human behavior data.The recognized human behavior includes daily behaviors(walking,upstairs,downstairs,sitting,jogging)and abnormal behavior(falling).It has important application value to provide scientific and effective exercise guidance for human health by using machine learning technology to analyze human motion data.Therefore,the main research work of this thesis is as follows:At first,we summarized and analyzed the current research status of human behavior recognition.After comparative analysis,the open machine learning database(http://archive.ics.uci.edu/ml/)of university of California,Irvine was selected for human behavior recognition.Then the acquired behavior data was processed by filtering,acceleration decomposition,adding windows and data normalization.It was convenient for subsequent feature extraction.Secondly,we extracted motion features from processed data and then reduced dimension of these features.The features were analyzed by statistical methods in time domain and frequency domain.In view of the problem that positive and negative areas cancel each other in the integration process of the grey absolute correlation degree model,this thesis improves the grey absolute correlation model to reduce dimension from the angle of relative change area,and the improved model effectively improves the reliability of feature selection.At the same time,classification algorithm was designed.By analyzing the characteristics of human daily behaviors and abnormal behavior,we first recognized the fall behavior,and designed a threshold-based classification algorithm and four classification algorithms based on machine learning.The classification effect of different classifiers was compared through the evaluation index of the algorithm.Aiming at two kinds of behavior which were difficult to distinguish between upstairs and downstairs,two new features were used to further distinguish.The use of new features improved the recognition effect of upstairs and downstairs.Then,strong classifiers were designed.Considering the poor classification effect of single classifier based on machine learning algorithm,this thesis proposed using AdaBoost algorithm to train a single weak classifier,so that it could eventually be trained to a strong classifier through continuous iterations to improve the classification effect.Finally,classification model was optimized.In view of the shortcomings of AdaBoost in multi-classification of daily behavior,the convolutional neural network was used to improve the classification effect of daily behavior by combining convolutional neural network with support vector machine.