The Research On Human Activity Recognition Based On Ensemble Model

Human activity recognition is a significant research field in machine learning.Human activity recognition is divided into two field:based on visual images and based on sensors.With the development of sensor technology,the rapid development of mobile Internet and the maturity of machine learning theory,human activity recognition based on mobile device sensors has been paid more and more attention.Mobile devices are portable,real-time,flexible and other advantages,and has been generally used in human activity recognition.Human activity recognition is widely used and has commercial value,the specific application scenarios such as:human-computer interaction,motion aids,augmented reality,intelligent monitoring and so on.Human activity recognition is a classification problem.Therefore,the accuracy of classification results is particularly important.If the accuracy of recognition can not reach the acceptable range of applications,it may cause a negative impact on the user.The accuracy of recognition is base on the algorithm model.The quality of a model can impact the final result greatly.Therefore,by experiments,this thesis come to a conclusion that problems of non-linear classification and confusable samples will affect the accuracy of recognition.Then,this thesis study ensemble model and model parameter selection to improve accuracy of recognition from aspects of model optimization.In this thesis,the acceleration sensor data of a variety of human behavior are collected under mobile intelligent devices.For the raw data,we do data preprocessing and feature extraction to get the data samples which can be used in machine learning model training.Compared with the DTW template matching algorithm,the machine learning classification models have higher recognition accuracy.In this thesis,common machine learning models are used for modeling and analyzing the data set.The results show that the common machine learning models can get a good recognition accuracy.Nevertheless,the traditional machine learning models have a lot of room for improvement.The experimental results show that,although the data has been preprocessing,but there still exist the problems of non-linear classification and confusable samples.Individual models are often difficult to deal with those problem.By analyzing the advantages and disadvantages of common machine learning models,this thesis proposes two ensemble models based on the optimization of feature space and model classification.Compared with a single model,the new model can solve the above two problems effectively and improve the accuracy of recognition.Finally,the model is validated by experiments,and the parameters of the model are studied.