Human Action Classification using seismic sensor and machine learning techniques

• Published in: 2021 5th International Conference on Information Systems and Computer Networks (ISCON) pp. 1 - 6
• Main Authors: Chakraborty, Mainak, A, Srinivasan, Reddy, Srinivasa, Kumar Mandal, Sanjib, Bhaumik, Subhasis
• Format: Conference Proceeding
• Language: English
• Published: IEEE 22-10-2021
• Subjects: classification; Data models; Feature extraction; Human activity classification; machine learning; Machine learning algorithms; Real-time systems; seismic sensor; Surveillance; Training; Vibrations
• DOI: 10.1109/ISCON52037.2021.9702317

In this work, we propose a novel method for identifying human activities like walking and running, utilising ground vibration obtained from seismic sensor and using machine learning techniques. The proposed methodology is based on statistical feature extraction. It is grounded on the idea, that each of the activity generate distinctly unique seismic signatures. We curated a unique dataset of various activites using off-the-shelve geophones and 16 bit analog to digital converter. The seismic data were sampled at 10000Hz. The datasets we recorded is 2 min long each. Single and 6-channels sensors are used to record the dataset. The data recorded is denoised using bandpass filter. The denoised data is used to detect peaks by comparing with neighbouring values. Peak based segementation is done and various statistical features are extracted from the dataset. The effectiveness of feature extraction is increased by converting the data in spectral domain, by extracting the power spectrum. The feature extracted are labelled and is used to train machine learning models. We have explored different machine learning algorithms and Random Forest algorithm gives the accuracy of 91.62%.