New human action recognition scheme with geometrical feature representation and invariant discretization for video surveillance

Abstract

Human action recognition is an active research area in computer vision because of its immense application in the field of video surveillance, video retrieval, security systems, video indexing and human computer interaction. Action recognition is classified as the time varying feature data generated by human under different viewpoint that aims to build mapping between dynamic image information and semantic understanding. Although a great deal of progress has been made in recognition of human actions during last two decades, few proposed approaches in literature are reported. This leads to a need for much research works to be conducted in addressing on going challenges leading to developing more efficient approaches to solve human action recognition. Feature extraction is the main tasks in action recognition that represents the core of any action recognition procedure. The process of feature extraction involves transforming the input data that describe the shape of a segmented silhouette of a moving person into the set of represented features of action poses. In video surveillance, global moment invariant based on Geometrical Moment Invariant (GMI) is widely used in human action recognition. However, there are many drawbacks of GMI such that it lack of granular interpretation of the invariants relative to the shape. Consequently, the representation of features has not been standardized. Hence, this study proposes a new scheme of human action recognition (HAR) with geometrical moment invariants for feature extraction and supervised invariant discretization in identifying actions uniqueness in video sequencing. The proposed scheme is tested using IXMAS dataset in video sequence that has non rigid nature of human poses that resulting from drastic illumination changes, changing in pose and erratic motion patterns. The invarianceness of the proposed scheme is validated based on the intra-class and inter-class analysis. The result of the proposed scheme yields better performance in action recognition compared to the conventional scheme with an average of more than 99% accuracy while preserving the shape of the human actions in video images.