Application And Research Of Multiple Classifier Fusion Algorithm In Activity Recognition

With the rapid development of smart phones with built-in high-performance sensors in recent years,human activity recognition based on smart phones has become a hot research topic for scholars at home and abroad.Through the acquisition of data from mobile phone built-in sensors,the current user activity condition can be get quickly and accurately,which has brought great convenience to people's life and has been widely used in intelligent home furnishing,medical care,activity monitoring and other fields.Meanwhile,how to efficiently make use of smart phones to identify human activity becomes a difficult problem,so most researchers improve the recognition accuracy by optimizing the feature data or improving the classifier algorithm.However,preferred feature can not fully reflect the differences between various activity patterns,and a classifier algorithm may not have a good classification effect on a variety of activity.So it will be easy to cause the low accuracy of some activities recognition,the increase of the misjudgment rate of the entire model and other problems.In order to improve the accuracy of human daily activity recognition based on smart phones,we use the technology about the multiple classifiers fusion in ensemble learning,and build a model by taking advantage of different classifier processing objects that to be complementary with each other in this paper.According to the performance of the classifier in activity recognition,it is divided into special base classifier combination and common base classifier combination,and those two fusion algorithms are put forward against the output of the above two base classifier combinations.(1)A fusion algorithm based on soft decision.This algorithm uses the posterior probability value of each activity category which is output by base classifier for making a comparison,and chooses the largest probability value as threshold of the base classifier.Then.the outputs of all base classifiers will be summed and averaged,therefore obtaining final classification result.The algorithm uses maximum method,sum method,multiplication method,average method and other fusion algorithms of the base classifier output with the type of soft decision,and fusion processing is used against the output of special base classifier combination in specific activity training.(2)A fusion algorithm based on hard decision.This algorithm is a fusion algorithm that improves traditional voting method about the output of the base classifier for the hard decision to fuse.There are different types of base classifier in the common base classifier combination,and it is very different in ability of recognition.We need to evaluate the reliability of each base classifier in advance,and keep the decision of the base classifiers with high reliability as the valid votes.If it reaches the specified proportion,the activity category will be the final output result of the model.Through applying two models of activity recognition established by two base classifiers combination and two corresponding fusion algorithms,the users use their own smart phones for experimental verification in the real environment.The experimental results show that those two models of activity recognition can effectively identify human activities.In addition,the two fusion algorithms put forward in the paper have higher identification accuracy,reliability and stability than traditional fusion algorithm in terms of building model of activity recognition,which fully demonstrates its feasibility.