A random walk through human behavior

Youssef Chahir (1) , Youssef Zinbi (1) , Mahmoud Ghoniem (1) , Abderrahim Elmoataz (1)

1 Equipe Image - Laboratoire GREYC - UMR6072

In many applications, such as the video monitoring, the archiving and the video indexing, it is significant to recognize the movements of the people to be able to interpret their behaviors. This recognition of activity requires the extraction of multiple data, the automatic interpretation of image sequences, and called upon techniques of video analysis and techniques of data analysis and data classification. A human action being strongly related to the movement, we propose in this paper, an approach for tracking people to form a volume in 3D space (2d t). This volume which represents a given action will be characterized by 3D geometrical moments which are invariants with the translation and the scaling. In this article, we present a new approach of people actions categorization based on the Markov random walks on graph. The basic idea is to regard the whole of the actions (videos) as a weighted graph G=(V,E). This graph is defined as a set of vertices V which are represented by 3D volumes of the action, and a set of edges E which represent the similarity between actions. This similarity will be calculated by an Euclidean distance between the vectors characteristic of the actions. Then, we will describe the implementation of our approach and we show results of validation on a corpus of actions which represent different actions of several people.

Mots clés: Random walk; Nystrom; human action; 3D geometrical moments; graph cut; video

Domaines: Traitement des images [eess.IV]

Youssef Chahir, Youssef Zinbi, Mahmoud Ghoniem, Abderrahim Elmoataz. A random walk through human behavior. IS&T / SPIE International Conference on Multimedia Content Access: Algorithms and Systems III, 2009, San Jose, United States. pp.725503-1 - 725503-10. ⟨hal-00815818⟩