Study On Human Activity Recognition Based On Smartphone's Sensors

With the development of society and the progress of science and technology,more and more people begin to pay attention to human health.Real-time,accurate and efficient human activity recognition can be widely used in personal and family health monitoring,hospital medical rehabilitation,human-computer interaction,virtual reality,dance collection,film and television production and gait recognition.In recent years,the growing popularity of smart phones has greatly facilitated people's daily life;a variety of powerful embedded sensors make smart phones a ubiquitous platform for data acquisition and analysis,which also provides great potential for efficient human activity recognition,so human activity recognition through smart phone sensors has become an important lesson.Research topic.In the daily use of smart phones,the location of smart phones is changeable,which leads to the uncertainty of the orientation of the three coordinate axes of the smart phone inertial sensor.Fixed position human activity recognition is widely used in the existing research.This method is suitable for wearable human activity recognition,but not for smart phone sensor human activity recognition.In order to solve the above problems,based on the related research in this field and smart phone sensors,this paper discusses the problem of human activity recognition without position constraints.In this research,the earth coordinate system is used as the reference coordinate system of human body attitude,and the rotation matrix is calculated by quaternion method,which fuses the sensor data of acceleration,gyroscope and magnetometer.The sensor data of carrier coordinate system is converted to the earth coordinate system,thus solving the problem that the three-axis orientation of sensor can not be determined in the daily use of smart phones.Based on the sensor data,the time series of the sensor is extended,and the related features of the attitude are extracted based on the time series.Aiming at the high-dimensional features,a filtering combined feature selection method is designed,which eliminates the irrelevant and redundant features and obtains the optimal feature subset of the attitude representation.Finally,a human activity recognition model is constructed based on the optimal feature subset.The main contents of this paper are as follows:(1)Reasonable human activity experiments were designed and implemented.Three smart phones were placed in the hands of the experimenters,jacket pockets and trousers pockets respectively.Sensor data of acceleration,gyroscope and magnetometer were collected from 12 college students in 7 kinds of daily activitys.According to the collected sensor data,the sliding mean filter is used to pre-process,and the noise and interference in the acquisition process are removed.(2)The rotation matrix between the carrier coordinate system and the earth coordinate system is solved by the quaternion method of sensor data fusion,and the sensor data in the carrier coordinate system is transformed into the earth coordinate system,which means that the reference system of human body attitude is transformed from the carrier coordinate system to the earth coordinate system,thus eliminating the influence of the uncertain orientation of the three axes of the sensor on human body attitude.By extending the time series,a total of 22 groups of sensor data time series related to human activity are formed.For these time series,the time window for each human activity discrimination is set to be 6 s and overlap 50%.Based on the time window of human activity,12 time-domain features and 12 frequency-domain features are extracted.Finally,in order to eliminate the dimension difference between features,all features are mapped by arc tangent to realize dimensionless feature.(3)Aiming at a large number of irrelevant and redundant features in highdimensional features,a combined filtering feature selection algorithm based on ReliefF and FCBF algorithm is designed.Firstly,ReliefF algorithm is used to calculate the weights between features and labels to remove the irrelevant features below the threshold,and then FCBF algorithm is used to calculate the correlation between features and remove the redundant features with strong correlation between features.Through the combined filtering feature selection algorithm,21 features representing human activity are selected,which reduces the feature dimension and the time of feature calculation.(4)Based on the optimal feature subset of feature screening,the optimal parameters of the xgboost algorithm are searched by ten fold crossover and grid optimization algorithm,and a human activity recognition model is constructed based on the optimal feature subset and parameters.The recognition rate of the human activity without position constraints can reach 95.65%.