The Research Of Human Activity Recognition Technology Based On Sensor And Convolutional Neural Network

Human activity contains a wealth of information,and accurate identification of activities can be widely applied in the fields of smart medical care,sports health care and human-computer interaction,etc.With the continuous development of microelectronics and sensor technology,sensor-based human activity recognition(HAR)has received more and more attention due to its excellent convenience and privacy.At the same time,with the development of deep learning technologies such as recurrent neural network(RNN)and convolutional neural network(CNN),a large number of studies have applied these methods to HAR tasks.The designed end-to-end system avoids conventional hand-crafted features and improves the accuracy of activity recognition.Although the sensor-based HAR technology has made great progress in recent years,it still faces many problems and challenges.For example,how to further improve the training speed of deep models so that real-time learning on wearable devices becomes possible;how to further improve the feature extraction capabilities of deep models under the control of model parameters;how to make better use of data from multi-source sensors in the coming era of Internet of Everything.In view of the above problems,this thesis mainly researches the following aspects:(1)The Gramian Angular Field algorithm is introduced into the sensor-based HAR task,so that the one-dimensional time series collected by the sensor can be converted into a two-dimensional image format through steps such as scaling,coordinate axis transformation and trigonometric function calculation.It makes CNN-based deep learning algorithms suitable for sensor-based HAR task,and avoid the problems of slow RNN training speed and high time complexity.(2)Combining the structure and advantages of CNN networks such as residual learning and dilated convolution,a new multi-dilated convolution kernel residual module is proposed,which can effectively improve the multi-scale feature extraction capability of the model,thereby improving the accuracy of activity recognition.Based on this module,a single sensor based HAR algorithm is proposed,and the proposed algorithm is analyzed experimentally on a public data set containing only accelerometer.(3)Based on the proposed HAR algorithm with single sensor,the fusion of multi-source sensors is studied,and an automatic fusion network of multi-source sensors is designed,which can automatically perform feature layer fusion on data from different sensors.Comparative expperiments are performed on two public datasets containing multiple sensors to evaluate the performance of the proposed method.In this thesis,comparative experiments are performed on three public datasets including WISDM,UCI HAR and OPPORTUNITY.The results prove that the proposed feature extraction module and multi-source sensor automatic fusion network are effective,and can further improve the accuracy of activity recognition.The methods proposed in this thesis have broad application prospects and can be extended to many fields such as daily activities detection,smart medical care and smart home,etc.