Human Activity Recognition Using CNN Based On Embedded Sensors In Smartphones

With the progress of science and technology and the development of the times,smartphones have become an indispensable part of people's life.The needs of human for life have evolved from simple food and clothing problems to complex material,cultural and spiritual satisfaction.So,the smartphone-based human activity recognition(HAR)have become a hot research issue.Meanwhile,the HAR based on smartphone sensors provides an efficient way for studying the connection between human physical activities and health issues.In the paper,two feature sets are involved,including tri-axial angular velocity data collected from gyroscope sensor and tri-axial total acceleration data collected from accelerometer sensor.The two feature sets and their many kinds of transformations are used to divide activities into six types of activities like walking,walking upstairs,walking downstairs,sitting,standing and lying.We design three experimental schemes to study and improve the accuracy of human activity recognition.In the experimental scheme 1,two kinds of CNN architectures are designed for HAR.The one is Architecture A in which only one set of features is combined at the first convolution layer,and the other one is Architecture B in which two sets of the features are combined at the first convolution layer.The validation data set is used to automatically determine the iteration number during the training process.It is shown that the performance of Architecture B is better compared to Architecture A.And the Architecture B is further improved by varying the number of the features maps at each convolution layer and the one producing the best result is selected.Compared with five other HAR methods using CNN,the proposed method could achieve a better recognition accuracy of 97.5% for a UCI HAR dataset.In addition,both the experimental scheme 2 and 3 achieve the goals which have the inspiring results.In three experiments,experimental scheme 2 constructs a 3-D data signal input which is similar to "R,G and B" color channels,and its accuracy rate reaches 95%.In experimental scheme 3,we try to modify the convolution process of the first convolutional layer,so that each column of the generated feature maps come from different convolution kernels.Finally,the results prove that the scheme 3 achieves the expected results.