Activity Recognition in Surfing - A Comparative Study between Hidden Markov Model and Support Vector Machine

• Published in: Procedia engineering Vol. 147; pp. 912 - 917
• Main Authors: Hoettinger, Hannes, Mally, Franziska, Sabo, Anton
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 2016
• Subjects: Activity Recognition; Hidden Markov Model (HMM); Inertial Measurement Unit (IMU); Principal Component Analysis (PCA); Support Vector Machine (SVM); Activity Recognition; Hidden Markov Model (HMM); Principal Component Analysis (PCA); Support Vector Machine (SVM); Inertial Measurement Unit (IMU)
• ISSN: 1877-7058; 1877-7058
• DOI: 10.1016/j.proeng.2016.06.279

The present project describes a comparative study between two different machine learning approaches, the Hidden Markov Model (HMM) and Support Vector Machines (SVMs), for activity recognition in surfing, aiming to distinguish surfing from other non- traditional (non-surfing) movements. The Hidden Markov Model has been introduced as a probabilistic or statistical framework for time-varying processes, whereas the Support Vector Machine algorithm is probably the most widely used kernel learning algorithm. Human activities are classified by using only one Inertial Measurement Unit (IMU) worn on the chest. A feature set extracted from the raw sensor data is used in the classification process. Feature transformation, in respect of dimensional reduction is implemented with Principal Component Analysis (PCA). A performance comparison of the classification models is provided in terms of their correct differentiation rates and confusion matrices, as well as their preprocessing and training requirements. 5-fold cross validation is employed to validate the classifiers. The results indicate that the HMM results in a higher classification accuracy of 91.4% compared to the SVM with an accuracy of 83.4%. The algorithm is capable of classifying time-varying motions from input data of an IMU worn during a surfing session. Moreover, the surfing style between subjects differs widely from left to right waves, right to left waves, goofy or regular footed and the execution itself. However, the implementation of the wave-model allows to train only one data set including every wave data collected and must not separate the data into different forms of execution.