A human action can be identified by visualizing the sequence of 2D binary projections over time. Here, one of the most representative features is shape and a wide range of algorithms have been proposed using its descriptors. This paper proposes convex deficiencies, the difference between an object and its convex hull, to be considered as a representation for the human action classification problem. A simple description using the centroids of the convex deficiencies over time is presented. Recognition of human actions is done with a fast matching algorithm that considers the spatial distribution of the centroid trajectories and the shape of the clusters in its 2D projection. The proposed representation is robust to deformations, scale, speed of the performed action and to the starting point of the movement sequence. Experiments using the videos of the Weizmann database show promising results demonstrating the effectiveness of the proposed methodology in classifying simple human actions, e.g. walking and running. The new proposed methodology should be extendable to a broader set of actions.


Iglesias-Ham, M., Bartolo García-Reyes, E., Kropatsch, W., & Artner, N. (2011). Convex Deficiencies for Human Action Recognition. Journal of Intelligent and Robotic Systems, 64(3), 353–364. http://hdl.handle.net/20.500.12708/162740