Abstract
Context-awareness is an essential part of pervasive computing. Video-based human activity recognition (HAR) has arisen as an imperative module to detect user's situation for involuntary facility delivery in context-aware domains. The activity recognition systems are frequently employed for protective and practical health care. Most of the existing works utilize RGB (red, green, and blue) cameras which present confidentiality and security concerns in the health-care domain. The existing approaches also do not sustain their performance results under the presence of a depth camera. Moreover, the accuracy of an HAR system relies on the extraction and selection of the prominent features from the feature space. To address these limitations, in this research, we first employ a depth camera to resolve the confidentiality and security concerns, and propose an unsupervised segmentation algorithm that can accurately segment the human body from the video frame. Then, we propose a new feature selection technique that has the ability to excerpt and select the best features from the feature space. Our proposed feature selection method focuses on the selection of confined features from the series of images and discriminate their category based on reversion (i.e., regression) value. The proposed method extracts and selects the best features through the benefits of forward selection and backward regression algorithms. Finally, we have trained and tested our proposed system by employing hidden Markov model (HMM) to label the activities. The proposed approach presents a significant performance against the existing works using depth camera.