Human Activity Recognition Based On Tri-axis Accelerometer

With the development of sensor technology,the volume of sensors is getting smaller and smaller,and the portability is getting higher and higher.Sensor-based human activity recognition has attracted the attention of many researchers.At present,human activity recognition is mainly realized by machine learning methods and deep learning methods.High precision,high efficiency,specific application scenarios,multi-sensor fusion and other factors are important goals for solving human activity recognition tasks,but the hidden relationship between the internal axes of the sensors is less concerned.At present,the research methods for the hidden relationship between the internal axes of the sensors mostly rely on human domain knowledge and have certain limitations.In order to solve this problem and explore the influence of hidden relationship on classification accuracy,this paper proposes a convolutional neural network method based on multi-channel data fusion and a convolutional neural network method based on single-channel data fusion.The main work of this paper is as follows:(1)In order to utilize the hidden relationship between the internal axes of the sensor,this paper proposes to use the data fusion method to complete the acquisition of the fusion data including the hidden relationship between the internal axes of the sensor.After that,the convolutional neural network is used for feature extraction and activity classification.Since the original sensor data is 1D data and has multiple channels,this paper proposes a convolutional neural network method based on multi-channel data fusion according to its data format.(2)Considering that the 1D convolution kernel can only capture the local dependence of time,the 2D convolution kernel can capture not only the local dependence of time but also the local dependence of space compared with the 1D convolution kernel.On this basis,this paper proposes a convolutional neural network method based on single channel data fusion.The method converts 1D sensor data into 2D data(each row in the data is fused data)and inputs it into the convolutional neural network to improve the accuracy of activity recognition.According to the different sensor data formats,this paper proposes a multi-channel data fusion method and a single-channel data fusion method.Based on the research of human activity recognition,a multi-channel data fusion based convolutional neural network method and a single channel are proposed.Convolutional neural network method for data fusion.(3)In order to verify the human activity recognition method based on the tri-axis accelerometer proposed in this paper,the WISDM public dataset is used for experiments.Through a large number of experimental analysis,the effectiveness of the two data fusion methods is proved.The comparison between the two activity recognition methods and the convolutional neural network method proves that the activity recognition method using data fusion can obtain higher classification accuracy,and the convolutional neural network method based on single channel data fusion has better effect.Its accuracy rate reached 98.83%.The activity recognition method proposed in this paper effectively solves the problems existing in sensor-based human activity recognition research at present,and the method is more suitable for solving the problem of activity recognition with large difference in triaxial data.