Research On Human Activity Recognition Method Based On Motion Features

Human activity recognition,as a key research area in Human-Computer Interaction and ubiquitous computing,provides information on a user’s physical activities that allows computer to actively assist users with their tasks.Research in computer vision has been at the forefront of this work.The development of sensor technology caused a shift toward using inertial sensors,such as accelerometers or gyroscopes.Therefore,human activity recognition method based on wireless sensors is becoming a research hotspot.Feature extraction is one of the key problems of human activity recognition.Extracting the most effective feature is an important way to improve human activity recognition method.Existing sensor-based human activity recognition methods often extractfeatures from discrete data in the feature extraction stage,like mean,variance,and kurtosis of acceleration or angular velocity data.However,on one hand,these features form discrete data do not reflect the continuity of human motion.On the other hand,extracting handcraft features relies on the priori knowledge.Therefore,we mainly focuses on above shortages and propose our methods based on motion features with wearable motion capture system.First of all,limb movement is continuous in the process of human motion while the corresponding motion capture data is discrete.In Chapter 2,this dissertationproposes a method for human daily action recognition based on Functional Data Analysis(FDA)so as to analysis the continuity and periodicity of daily activity.Firstly,we transform the periodic data collected by the wearable motion capture system into functional data using FDA,and then define the continuity and periodicity of data exactly by using function properties.Then,the periodic data features and functional features are extracted from the fitting function respectively.Combine with the support vector machine algorithm,we consider them as the feature vector and effectively recognize multi-class human daily activities.Secondly,deep neural network can automatically learn the low-level and high-level distributed features fromdataso thatit can effectively replace the traditional handcraft features with self-learning features.In this dissertation,the concept of short-time activity is defined in Chapter 3,and a method of human activity recognition based on deep learning is proposed based on short-time activity.In thefirst stage,we construct an over-complete pattern library which includes different patterns of short-time human activity.This library is produced by segmenting a long-time activity with sliding window method.In thesecond stage,we extract self-learning features based onthe over-completed pattern library and convolutional neural network to realize the classification of human daily short-time activities.