Research On Feature Extraction And Fast Classification Algorithm Based On Wearable Devices

With the development of science and technology,the wearable devices have been rapidly popularized.Researchs based on wearable devices have also become a hot field,mainly concluding data mining and data security based on wearable sensor.This paper mainly studies the rapid recognition of human activity based on wearable devices,and the authentication based on the action of walking.At present,in the daily activity recognition,most methods achieve a high precision accuracy by increasing memory and time consumption or need a complex dataset with multiple sensors.In order to improve the recognition accuracy and speed,HL-HAR algorithm is proposed in this paper.This algorithm mainly adopts two important methods.Using LRS feature selection algorithm to reduce the dimension of the features and H-ELM algorithm to construct the classifier.The experimental results show that the LRS can remove redundant features,and select a feature vector which is most conducive to classifier for classification.H-ELM algorithm is a multi-layer structure of ELM algorithm.Comparing the ELM,the classification effect of the H-ELM is improved both in accuracy and in time,so that the memory consumption problem in the calculation process can be solved.Human Activity Recognition Using Smartphones Datasets is used in the experiment,the daily activity recognition is carried out by HL-HAR,the final selection of 25-dimensional feature vectors can get an overall accuracy of 93.7%.In the identification and authentication,we only extract the gait data from the activity of walking.As the walking process is a cyclic process,the pre-processed data is segmented according to the cycle.After the segmentation is done in a combination of maximum and minimum values,the data segments are normalized to a uniform length for experimentation.In this paper,DTW algorithm and SVM classification algorithm are used to carry out experiments respectively.Comparing the DTW algorithm with the traditional algorithm,the DTW algorithm has a higher recognition accuracy,which is 83.29%.When using SVM algorithm,the feature vector consists time-domain features and frequency-domain features are sent to SVM classifier.The time domain features include mean and variance,and the frequency domain featrues include the maximum amplitude in a certain frequency range.The recognition accuracy achieved by this method is 98.99%.In the authentication process,the DTW algorithm has an EER of 14.42% and an equal error rate of each object ranging from 6.12% to 34.62%.The verification result of the SVM algorithm is also good.