A Lightweight Human Activity Recognition Method Based On Stochastic Configuration Networks

With the implementation of the made in China 2025 strategy,all walks of life have ushered in huge development opportunities.Especially in the field of artificial intelligence.Human activity recognition,as one of the research hotspots in the field of artificial intelligence,has played an important role in many fields such as intelligent buildings,medical care,security,leisure and entertainment,and military.It consists of three steps.Firstly,the human activity data is collected by sensors,then the data is processed by feature engineering.Finally,various recognition models are used to model and classify the activity.There are two main methods for human activity recognition:video based recognition and wearable sensor based recognition.The research of human activity recognition based on video image mainly depends on the camera,so it is vulnerable to privacy,lighting and background problems.Human activity recognition based on sensors,especially based on the built-in sensors of smart phones,has attracted more and more researchers' attention due to its non-invasive,universality,small size and low cost.Although the research of human activity recognition has made a lot of remarkable achievements,due to the limitation of smart phones resources(CPU,storage space),it is urgent to propose a human activity recognition model with fast modeling speed,high precision,compact structure and high lightweight.Based on the in-depth analysis of researches on human activity recognition at home and abroad,this paper conducts research on human activity recognition in two aspects: feature engineering and activity recognition models:(1)Propose a feature selection method combining Principal Component Analysis(PCA)and Neighbourhood Components Analysis(NCA).For machine learning,the quality of the input data directly determines the upper limit of the model.Correspondingly,in human activity recognition,the quality of features selected by feature engineering also has an important impact on the classification performance of activity recognition models.Aiming at the problem of feature selection in feature engineering that only focuses on feature redundancy or correlation,this paper proposes a feature selection method based on the fusion of principal component analysis and nearest neighbor component analysis.This method can reduce feature redundancy.At the same time,it improves the relevance of features,thereby increasing the separability of activity feature sets.The experimental results show that the features obtained by this method are helpful to improve the recognition accuracy and modeling speed of the model.(2)Propose a stochastic configuration network(SCNs)human activity recognition algorithm based on manifold regularization.In view of the characteristics of SCNs that stochasticly allocate hidden layer node parameters in the dynamic interval according to the supervision mechanism,it increases the modeling speed and also causes the output space of the data hidden layer to show a certain non-linear distribution.While the global least square method used in the output layer is unable to mine this non-linear structure,which leads to the problem of not compacting the constructed model.In this paper,the manifold regularization method that can keep the internal structure of the data unchanged during projection transformation is used to improve the problem of stochastic configuration of the network,and then improve the compactness of the stochastic configuration of the network structure.Finally,the proposed stochastic configuration network algorithm of manifold regularization is applied to the research of human activity recognition.The simulation experiment results show that the manifold regularization method helps to improve the compactness of the stochastic configuration network.(3)Propose a stochastic configuration network human activity recognition algorithm based on dynamic and stepwise updating of output weights.Aiming at the problem of time-consuming calculation of output weights in the stochastic configuration network based on manifold regularization and the original stochastic configuration network.In this paper,the dynamic step-by-step update method is used to replace the time-consuming Moore-Penrose generalized inverse in the original method to calculate the output weight of the network.By improving the modeling speed of the stochastic configuration network algorithm,the computational consumption of computing equipment is reduced,and the lightness of the algorithm is improved.The simulation experiment results show that dynamically updating the output weights step by step helps to improve the lightness of the stochastic configuration network.To sum up,this article focuses on the existing research problems of human activity recognition,and conducts research from two aspects: feature selection and classification model.First,a feature selection method based on the fusion of PCA and NCA is proposed;secondly,a structured compact human activity classification model and a lightweight human activity recognition algorithm are proposed based on a stochastic configuration network.To sum up,aiming at the existing research problems of human activity recognition,this paper studies from two aspects: feature selection and activity recognition model.Firstly,a feature selection method based on PCA and NCA fusion is proposed;secondly,a compact human activity recognition algorithm and a lightweight human activity recognition algorithm are proposed based on stochastic configuration network.This thesis includes 17 figures,7 tables and 75 references.