A Research Of Sparse Representation Recognition For Action Pattern In Body Sensor Networks

In recent years,emerging body area network technologies have received extensive attention and applications in health monitoring,film and computer game production and professional motion analysis.Most of current related research attempts to solve the problem of action recognition in body sensor networks by using sparse representation classification algorithm.However,there are some problems in SRC algorithm,including the oversized scale of over-complete dictionary and the overmuch error of sparse representation coefficients.In this article,we attempt to save the problems above by using dictionary training algorithms,(joint)sparse representation classification algorithm,collaborative representation classification algorithm and so on.The main researching work in this article is following:1.The over-complete dictionaries in traditional sparse representation classification algorithms consists of a large number of high-dimension action signals,which results in the oversized scale of dictionaries and influences action recognition algorithms'performance.The sparse representation coefficients residuals are obtained by simple average superposition and it is difficult to approach the distribution of different modes of actions in space.Designing for the issues above,a maximum likelihood sparse representation action classification algorithm based on K-SVD in body sensor networks is proposed.In this algorithm,all of action training samples are grouped according their classes to be trained,respectively.The mutual interference among different groups in the process of training can be avoided and sub-dictionaries for every class can be obtained.Then,these sub-dictionaries are used to construct an over-complete dictionary.And the dictionary is able to sparsely represent the testing samples precisely.The sparse representation coefficients are precisely approximated by maximum likelihood sparse model and the recognition results are determined by the coefficients.The experimental results show that the accuracy of the proposed algorithm is clearly better than traditional sparse representation algorithm.It is able to efficiently improve action recognition based on Body Sensor Networks.2.In the process of multi-sensor action signals dealt by traditional sparse representation classification algorithms,the signals are usually regarded as mutually independent variables,which results in that the associated characteristics among action signals from different sensors can't be completely dig out and effectively utilized.In this article,joint sparse representation classification algorithm is adopted to process multi-sensors action signals and D-KSVD algorithm has be improved to design the over-complete dictionary.Therefore,a joint sparse representation action recognition algorithm based on D-KSVD is proposed.Based on modified D-KSVD algorithm,the sub-dictionaries are optimized respectively,which is able to avoid the mutual interference among different kinds of action and promote the discriminative ability of dictionary.Then,the sparse representation coefficients of the testing sample can be obtained by joint construction method and the recognition results are determined by these sparse coefficients.The experimental results show that the proposed algorithm is able to extract the spatio-temporal correlation among the action signals recorded by multi-sensor.The proposed algorithm can largely improve the accuracy of action signals based on multi-sensor Body Area Networks.3.The traditional sparse representation classification algorithm is used to identify multi-sensor motion data with low efficiency.In order to reduce the time required for action recognition,the nearest neighbor idea is utilized to greatly reduce the dictionary size.The idea of collaborative representation is also merged in the proposed algorithm and a fast robust collaborative representation action recognition algorithm is proposed in this article.This proposed algorithm attempts to search the neighbor classes and samples of a test sample based on the nearest neighbor principle,which is helpful to decrease the calculation cost of the classification algorithm.Then robust collaborative representation classification model is built.The representation coefficients and residuals can be obtained by solving Augmented Lagrange Multiplier algorithm.The labels of test samples are determined by the representation residuals.The experimental results show that the proposed algorithm is able to acquire more correlation and coordination from multi-sensor action signals.The proposed algorithm can extremely decrease the time complexity of action recognition algorithm and efficiently increase the accuracy of action models at the same time.The wearable action recognition database is adopted to be the experimental data resource.The proposed algorithms are estimated by contrastive experiments from multi-reviews,including action recognition rate and time complexity.The experimental results show that the proposed algorithms overplay traditional machining learning algorithms and sparse representation classification algorithms and are able to improve the classification efficiency of human body action recognition based on BSN.This research provides new ideas and methods for the continuous related researching works.