Action Recognition Using Action Unit Based Feature Extraction and SURF Descriptor

A model is proposed for human action recognition using a new concept of action units to represent human actions in videos. Two stages are involved in our approach. First, training stage learns the model for action units and action classifiers. Second, testing stage uses the learned model for action recognition. Different interconnected components are used for action recognition. First involves the use of a new descriptor named SURF which is the fastest one with good performance and shows its advantages in rotation, blur and illumination changes. SURF uses square-shaped filters which forms an approximation of Gaussian smoothing. Second, involves learning action units using a factorizing algorithm called the graph regularized nonnegative matrix factorization, which helps to encode geometrical information. Third, a model is proposed to select the discriminative action units for better action recognition. SVM model is adopted as the predictive model.