Human Activities Recognition using Semi-Supervised SVM and Hidden Markov Models

Abstract

Automatic human activity recognition is an area of interest for developing health, security, and sports applications. Currently, it is necessary to develop methods that facilitate the training process and reduce the costs of this process. This paper explores a methodology to classify human physical activities in a semi-supervised paradigm. With this approach, it is possible to reduce the number of labels necessary to train the learning model and the complexity of this process. This process begins by deducting the number of micro-movements or sub-movements where the data should be grouped and assigning the label through a clustering technique. We perform this procedure for a specific group of micro-movements whose label is unknown. Later, the classification process starts by using two methods, a Support Vector Machine (SVM) that identifies the micro-movements and a Markov Hidden Model that detects the human physical activity as a function of sequences. The results show that with a percentage of 80 % of the known labels, we achieved outcomes like the supervised paradigms found in the literature. This facilitates training these learning models by reducing the number of examples requiring labels and reduces the economic costs, which is one of the significant limitations of machine learning processes.

Keywords

• spectral clustering
• semi-supervised learning
• motion estimation
• data fusion
• human activity recognition