Machine Vision for Human Activity Recognition: Features & Algorithms

Human Activity Recognition (HAR) is a multifaceted aspect of computer vision and machine learning, which encompasses group activity pattern discovery, interpersonal interaction analysis, human gesture and action recognition. It has proliferating demands from wide applications, such as visual surveillance and security, entertainment, healthcare systems, video indexing, human-computer interaction and video retrieval. So over the last decade, a diversity of approaches has been developed to investigate the HAR. We overcome their limitations by proposing new robust features and the algorithms to build the unified HAR framework. Features play a vital role in HAR. Global features are generated using the entire video sequence while ignoring explicit temporal information but they capture the oriented and holistic underlying patterns. We found that HAR can be improved by fusing extra temporal information with global representation.