Human activities recognition via features extraction from skeleton

Human activities recognition (HAR) enabling the understanding of basic human actions from still images has overriding importance in computer vision and pattern recognition for sundry applications. We propose a novel method for HAR by taking out the skeleton from the image for extracting useful features. This approach comprised of two steps namely (i) an automatic skeletal feature extraction and partitioning into two parts as block that determines angles between terminals and (ii) HAR by using non-linear Support Vector Machine (SVM). The model performance is evaluated using three available challenging datasets such as INRIA, KTH and Willow-action all with seven activities and each possessing eight scenarios. The images are normalized in (64×128) pixels format from digital silhouette via circle algorithm. Our method efficiently achieves a recognition rate as much as 86% with excellent features. The proposed model being highly promising compared to the existing one may contribute towards the development of computer vision architecture.