Silhouette-based action recognition using simple shape descriptors

Abstract

This paper presents human action recognition method based on silhouette sequences and simple shape descriptors. The proposed solution uses single scalar shape measures to represent each silhouette from an action sequence. Scalars are then combined into a vector that represents the entire sequence. In the following step, vectors are transformed into sequence representations and matched with the use of leave-one-out cross-validation technique and selected similarity or dissimilarity measure. Additionally, action sequences are pre-classified using the information about centroid trajectory into two subgroups—actions that are performed in place and actions during which a person moves in the frame. The average percentage accuracy is 80%—the result is very satisfactory taking into consideration the very small amount of data used. The paper provides information on the approach, some key definitions as well as experimental results.