Research On Energy Efficiency-Aware Human Activity Recognition

With the progress of science,technology and the development of intelligent devices,human activity recognition plays a huge role in many fields,including medical care,human-computer interaction,smart home,game entertainment,etc.Therefore,the research has received more and more attention.At the same time,because of the small size of smart phones,multi sensor integration and high computational efficiency,researchers begin to focus on activity recognition based on smart phones.As smartphones need to be able to be used in a sustainable way,the energy consumption of smart phones is also a limiting factor in the development of this field.Hence,this thesis mainly focuses on the two parts,one is the smartphone-based activity recognition and the other one is energysaving.In this way,this thesis can ensure the recognition accuracy of six basic actions and minimize the energy consumption.Aiming at the problem of insufficient recognition accuracy and high computational complexity in the process of activity recognition,this thesis selects machine learning methods in machine learning and deep learning methods,and at the same time,according to the principle of feature extraction and selection that can reduce computational complexity,the technique of linear discriminant analysis is also used before sending the raw data to the classifier.In addition,in order to optimize the classifier parameters as much as possible,grid search and cross-validation algorithms are also incorporated in the parameter adjustment process.Finally,for five different machine learning classifiers,the support vector machine with linear kernel is found to be the optimal classifier in this experiment by comparing and analyzing the recognition accuracy and computational time complexity.In addition to the problem of energy saving,according to the principle that the higher the sampling rate,the higher the energy consumption,this thesis first proposes an adaptive sampling adjustment algorithm from the data point of view,in which the smoothness index is innovatively proposed,and according to this,the sampling rate is dynamically adjusted to achieve the purpose of reducing energy consumption.When the data collection is completed,it is sent to the machine learning classifier to realize behavior recognition.Finally,it is found that the algorithm in this thesis can reduce energy consumption by56.39% while at the same time maintain the recognition accuracy rate at 99.58%.In view of energy consumption,if the sampling rate is adjusted from the perspective of the data itself,the sampling rate cannot be adjusted immediately at some behavior switching.Based on this principle,this thesis adopts an adaptive adjustment algorithm of sampling rate from the perspective of behavior types.First,the linear discriminant analysis combined with the support vector machine classifier is used to predict the behavior at the current sampling time,and compare it with the behavior at the previous sampling time.Through multiple experiments,the best sampling adjustment time and the best sampling rate for different activities are selected.Finally,the algorithm is verified on the UCI-HAR dataset.It is found that the algorithm in this thesis can flexibly switch the sampling rate of smartphones with saving 60.77% of energy consumption,and also achieving a recognition accuracy of 98.62%.