Robust Spatio-Temporal Features for Human Action Recognition

In this chapter we describe and evaluate two recent feature detectors and descriptors used in the context of action recognition: 3D SIFT and 3D SURF. We first give an introduction to the algorithms in the 2D domain, named SIFT and SURF. For each method, an explanation of the theory upon which they are based is given and a comparison of the different approaches is shown. Then, we describe the extension of the 2D methods SIFT and SURF into the temporal domain, known as 3D SIFT and 3D SURF. The similarities and differences for both methods are emphasized. As a comparison of the 3D methods, we evaluate the performance of 3D SURF and 3D SIFT in the field of Human Action Recognition. Our results have shown similar accuracy performance, but a greater efficiency for 3D SURF approach compared with 3D SIFT.