Space-time Templates based Features for Patient Activity Recognition

Human activity recognition has been the popular area of research among the computer vision researchers. The proposed work is focused on patient activity recognition in hospital room environment. We have investigated the optimum supportive features for the patient activity recognition problem. Exploiting the strength of space-time template approaches for activity analysis, Motion-Density Image (MDI) is proposed for patient's activities when used supportively with Motion-History Image (MHI). The final feature vector is created by combining the description of MHI and MDI using Motion Orientation Histograms (MOH) and then applying Linear Discriminant Analysis (LDA) for dimensionality reduction. The LDA technique not only reduced the complexity cost required for classification but also played vital role to get best recognition results by increasing between-class separation and decreasing the with-in class variance. To validate the proposed approach, we recorded a video dataset containing 8 activities of patients performed in hospital room environment under indoor conditions. We have successfully validated the results of the proposed approach on our dataset by training the SVM classifier and achieved 97.9% average recognition accuracy.