Human Activity Recognition Based on Spatial Distribution of Gradients at Sublevels of Average Energy Silhouette Images

The aim of this paper is to present a unified framework for human action and activity recognition by analysing the effect of computation of spatial distribution of gradients (SDGs) on average energy silhouette images (AESIs). Based on the analysis of SDGs computation at various decomposition levels, an effective approach to compute the SDGs is developed. The AESI is constructed for the representation of the shape of action and activity and these are the reflection of 3-D pose into 2-D pose. To describe the AESIs, the SDGs at various sublevels and sum of the directional pixels (SDPs) variations is computed. The temporal content of the activity is computed through R-transform (RT). Finally, the shape computed through SDGs and SDPs, and temporal evidences through RT of the human body is fused together at the recognition stage, which results in a new powerful unified feature map model. The performance of the proposed framework is evaluated on three different publicly available datasets, i.e., Weizmann, KTH, and Ballet and the recognition accuracy is computed using hybrid classifier. The highest recognition accuracy achieved on these datasets is compared with the similar state-of-the-art techniques and demonstrate the superior performance.