Human action recognition via affine moment invariants

Despite their high stability and compactness, affine moment invariants have received a relatively little attention in action recognition literature. In this paper, we introduce an approach for activity recognition, based on affine moment invariants. In the proposed approach, a compact computationally-efficient shape descriptor is developed by using affine moment invariants. Affine moment invariants are derived from the 3D spatio-temporal action volume and the average image created from the 3D volume, and classified by an SVM classifier. On KTH dataset, the method achieves performance results that compare favorably with these of other contemporary approaches reported in literature.