Recognition System Of Human Activities Based On Time-Frequency Features Of Accelerometer Data

With the development of Micro-Electro-Mechanical System, wearable sensor-based human activity recognition systems have important applications in various fields such as health management, motion analysis, military and industry. In this paper, we propose a time-frequency features extraction method based on wavelet transform, which extracts 5 time-frequency features, namely wavelet entropy, wavelet energy, wavelet waveform length, wavelet coefficient variance and wavelet coefficient standard deviation. The experimental results are evaluated on the publicly available benchmark WISDM dataset including accelerometer data. Our model achieves 99.2%, 99.1% and 95.6% test accuracy on Subspace KNN, Bagged tree and Gaussian SVM respectively.