Research On Human Activity Recognition Algorithms Based On Mobile Phone Sensors

With the development of artificial intelligence,human activity recognition technology has gained widely attention in recent years.At present,collecting human activity data by mobile phone's built-in sensors has become a new direction of the research on human activity recognition.However,compared with the traditional wearable data capture system,the location and direction of the phone is not fixed,which is more difficult to recognize human activities.In this paper,the human activity recognition algorithms are studied and simulated while the human activity data is collected by mobile sensors,and the main contents of the algorithms include the original data preprocessing,feature extraction and selection,classification,etc.The author's main research works are as follows:(1)The human activity data is collected by mobile sensors and data preprocessing process is designed.Four different phone placement locations are taken into account in the data collection process,and 26 kinds of human activity data are collected by a 3-axis accelerometer,a 3-aixs gyroscope and a 3-aixs magnetometer inside the phone,and activity categories include standing,walking,sitting down,standing up,falling,and so on.And the data preprocessing process,including signal windowing,filtering,acceleration signal decomposition and generating other signals,is used to remove the interference of mobile phone's orientation and location,and to provide more activity information.(2)An improved feature selection algorithm is proposed.After feature extraction from the preprocessing data,as for feature selection,the SFFS(Sequential Forward Feature Selection)algorithm is investigated and improved.According to the weaknesses of SFFS algorithm that it spends too long time on feature selection when the number of features is enormous and the selected features can not be removed,the feature weight in the ReliefF algorithm is introduced to remove most features in advance which have poor ability to recognize,and combining with SBFS(Sequential Backward Feature Selection)algorithm in the iterative search method to remove the redundancy in the selected features.The experimental results show that the improved feature selection algorithm proposed in this paper can improve the recognition performance of the corresponding human activity recognition algorithm.(3)An improved SOA-SVM classification algorithm is proposed.Firstly,SVM classification algorithm is researched and analyzed,and the model's parameters are optimized by comparison experiments to determine the multi-classification method and the kernel function of SVM model in human activity recognition algorithm.Next,the SOA optimization algorithm is researched,and improved by introducing a transfer factor based on the Metropolis criterion in the SA algorithm and a mutation factor to improve the optimal performance of the SOA algorithm and avoid involving local optimal solution,and the experimental results show that the improved SOA algorithm has better optimization performance.And then,the improved SOA algorithm is introduced into the SVM algorithm,in order words,the improved SOA algorithm is used to determine the optimal kernel function parameter and the optimal penalty factor of the SVM model,and the experimental results show that the improved SOA-SVM algorithm has better recognition performance in human activity recognition.(4)A hierarchical human activity recognition algorithm is designed and simulated.A multi-classifier combination model is designed to recognize the 26 kinds of human activities in collected data.Firstly,aiming at the different recognition requirements of each classifier in the multi-classifier combination model,the corresponding input feature vectors of all classifiers are obtained by the improved feature selection algorithm.Then,the hierarchical activity recognition is realized by using some improved SOA-SVM classifiers and the corresponding feature vector sample sets.The simulation results show that,for the 26 kinds of human activity data collected by mobile phone in different locations,the hierarchical human activity recognition algorithm based on improved feature selection algorithm and improved SOA-SVM classification algorithm can achieve 98.2% average recognition accuracy.