Research On Key Technologies Of Wearable Real Time Activity Recognition Based On Online Learning

Nowadays,with the increasing pressure of life and work,people pay more attention to their physical condition and want to know their physical and mental condition through their daily activities.Human behavior recognition plays an increasingly important role in people’s life,including medical,health,entertainment and monitoring.The contents are divided into four parts shown as follows.The traditional methods of human behavior recognition are often with the help of wearable devices,camera,recorder and other devices to obtain the original data,and then get the results by offline analysis.However,the results obtained by these methods are often lagging behind,the process of obtaining information and calibrating data is also very complicated.How to quickly and conveniently collect data for human behavior recognition has become a remarkable research direction.This paper proposes a semi supervised learning,which can effectively use the internal information of unlabeled data.For model data enhancement,Tri-training and VFDT are combined to improve accuracy.The proposed method is tested on different data sets.After inputting unlabeled data,the classification accuracy of the model is significantly improved;compared with the original semi supervised framework,much computing time and data storage space are saved.In addition,the model is reasonably simplified and embedded in wearable devices for online learning.The contents are divided into four parts shown as follows:1)The wearable social behavior recognition system is designed,and the signal collection experiment of college students is carried out.The behavior signals under different states of motion are collected,and the data features are extracted,and dataset is made.2)This paper proposes a collaborative learning method,which uses VFDT tree as the base classifier.Based on the characteristics of VFDT tree for streaming data,semi supervised framework Tri-training is improved.Compared with the original model Tritraining,the proposed method only needs a small amount of data to update the model.Histogram algorithm is used to change the way of data storage.Experiments on public and real datasets,the proposed method is compared with several similar tree models.The reasons for the different results are analyzed and the application scenarios of the model are discussed.3)The simulated annealing algorithm is used to filter the high-dimensional behavior features,and then the model is pretrained offline.After initialization,and the model is embedded on the wearable device.What’s more,the real-time recognition experiment based on wearable device is carried out.Testers wear wearable device continuously while switching different motion states during the experiment.The model is applied to the real data stream,learning online learning and real-time recognizing synchronously.4)The results of online activity recognition are compared with those of offline classification,the reasons are analyzed.