Activity Recognition Based On Deep Adaptive Incremental Learning

Human activity recognition(HAR)based on wearable devices and smartphones is becoming a revolutionary technology for personal health monitoring In the past decade,deep learning has been evolving to a prevalent method in the field of HAR due to its high accuracy and automatic feature extraction.However,big data based deep learning model has two limitations in HAR application: First,it is very time-consuming to label a large number of data samples;Secondly,when the training data and test data come from different sensor positions or users,the model performance will be significantly reduced.This paper addresses the transfer of activity recognition between different users(cross-person)and different positions(cross-position),explores how to build a more efficient HAR model by using source data with rich annotation and target domain data with just a few or no annotations.A hybrid model based on deep adaptive incremental learning is proposed,which combines the advantages of deep learning,transfer learning and online sequential extreme learning machine(OS-ELM):(1)A feature extractor based on convolutional neural network(CNN): The Squeeze and Excitation(SE)and global average pooling layer are added to standard CNN model.The improved model utilizes an end-to-end structure to automatically and efficiently extract advanced and meaningful features,adaptive to the input data size and learn the weight of different sensor feature channels;(2)A deep adaptation(DA)model for reducing domain shift: Two widely used statistical measurement methods,namely,maximum mean difference(MMD)and improved local mean difference(LMMD),as well as the anti-adaptation method of gradient reversion layer GRL,are employed to build a deep transfer adapdation model for reduce performance degradation due to the difference of data distribution between different users and different position sensors;(3)An adaptive classifier based on incremental learning of OS-ELM: OS-ELM can constantly update learning parameters with just a small amount of online updated labeled data in target domain,which has fast online learning speed and strong generalization ability.In this paper,six big public HAR datasets are selected to train and test the hybrid model.The experimental results show that when the user(person)or sensor position changes,the activity recognition accuracy is significantly better than the standard single CNN and deep transfer learning model.For example,in the cross-person experiments of seven common activities: the hybrid model can improve the recognition accuracy of m Health,DSADS and PAMAP2 datasets by more than 10%,and the final recognition accuracy of m Health and DSADS datasets are more than 95%,the PAMAP2 dataset is more than 75%;As for cross-position experiments: the recognition accuracy of m Health and PAMAP2 datasets are improved by more than 30%,DSADS dataset is improved by more than 20%,and the final recognition accuracy of m Health,DSADS and PAMAP2 datasets are respectively more than 85%,80% and 65%.