Department Of Computer Science, National Institute of Technology, Ichchhanath Surat - Dumas Rd, Gujarat 395007
Dr. Chetan Singh Thakur
Department Of Electronic Systems Engineering, Indian Institute of Science, Bangalore, CV Raman Rd, Bengaluru, Karnataka 560012
Inertial measurement units namely IMU sensors are widely used to gain data regarding human activities. These sensors provide us with the data regarding the motion of the body in the form of different sensor readings like accelerometer, gyroscope, magnetometer and more. This data is utilized in applications like determining direction in GPS system, head motion in AR/VR systems and even in medical applications to detect Parkinson’s disease by observing strange body motion of the affected person. This paper employs the use of raw data collected from 6 degree of freedom IMU sensor that comprises of accelerometer and gyroscope in order to predict certain human activities. There are in total 12 different physical human activities taken into account here. The raw data has been taken from University of California’s machine learning repository. Information regarding the raw data and experiment conditions can be acquired from their website. Feature extraction makes use of both frequency domain as well as time domain techniques over the given sampled raw data. While frequency domain features have been extracted from the spectrogram of raw data, time domain features include parameters like variance, kurtosis, skewedness, mean etc. Datasets having different number of features have been created and tests have been conducted on each of them separately. In order to overcome the problem of scarcity of data, data augmentation technique of down-sampling by local averaging and shuffling have been utilized. The generated datasets have been tested rigorously on a machine learning algorithm named support vector machine with a one vs rest approach. The results comprising of the training and test accuracy along with the confusion matrix have been provided.
Keywords: time domain, frequency domain, gyroscopes, accelerometers, physical activities, data augmentation