Human Activity Recognition Using Relative Relationship between Different Human Body Regions

The Human Activity Recognition (HAR) system can be used for various purposes, including video surveillance and Senior Care, by tracking daily activities. We propose the HAR system based on the relative relationship between several human body regions. The Region of Interest (ROI) is retrieved from the video signal of the subject conducting the activity from the area of the video frame. The ROI contains humans performing the action. The human body is divided into six different regions. The spatial-temporal features are extracted for each human region. As the spatial-temporal feature, Motion Vectors (MV) and Optical Flow (OF) are calculated, then quantized to form a histogram. A relative correlation is calculated between one body region and the remaining body region. This correlation is stored in a vector to create a feature vector for HAR. The dimension of the feature vectors is reduced using Principle Component Analysis (PCA) and Discrete Fourier Transform (DFT). The extracted feature vectors are used by Hidden Markov Model (HMM) to model human activities in the Dataset. For testing, the likelihood is calculated for the test signal. The class for the test signal will be the HMM model, for which we get the maximum likelihood. The HMM models are tested on KTH and SFB588 Basic Kitchen Tasks Dataset. The effect of feature selection methods on the system's overall accuracy is studied. An experimental result on the human activity datasets suggests that the proposed human activity model performs moderately well compared to the other human activity models.