Hindawi Publishing Corporation International Journal of Distributed Sensor Networks Volume 2013, Article ID 383906, 11 pages http://dx.doi.org/10.1155/2013/383906 Research Article A Feature Selection Approach to the Group Behavior Recognition Issue Using Static Context Information Alberto Pozo, Miguel A. Patricio, Jesús García, and José Manuel Molina University Carlos III of Madrid, 28270 Colmenarejo, Spain Correspondence should be addressed to Alberto Pozo; alberto@pozoesteban.es Received 18 February 2013; Revised 20 May 2013; Accepted 4 June 2013 Academic Editor: Juan Manuel Corchado Copyright © 2013 Alberto Pozo et al. Tis is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Tis paper deals with the problem of group behavior recognition. Our approach is to merge all the possible features of group behavior (individuals, groups, relationships between individuals, relationships between groups, etc.) with static context information relating to particular domains. All this information is represented as a set of features by classifcation algorithms. Tis is a very high- dimensional problem, with which classifcation algorithms are unable cope. For this reason, this paper also presents four feature selection alternatives: two wrappers and two flters. We present and compare the results of each method in the basketball domain. 1. Introduction Human or other types of behavior recognition are currently one of the most prolifc felds of research. One typical restriction in this feld is that the scene should contain only one element. Te system analyzes the behavior of the element and recognizes what the element is doing. Several papers address this aim [1, 2]. However, many situations, like team sports, animal social behavior (ethology), trafc analysis, and so forth, hinge on more than one element, where elements behave socially with respect to the other elements of the group. Individual behavior depends on group behavior in such cases. Terefore, individuals have to be studied together, as a group. Research on this area is scant. Some researchers deal with the elements as a crowd and try to recognize their behavior by analyzing shape or some other features. On the other hand, some authors deal with each element as an individual that is a member of the group. Our proposal takes this approach. Behavior recognition can be divided into two steps: feature extraction and feature analysis. In feature extraction, the system chooses and extracts the main raw information (from video cameras, GPS, RFID, TOF cameras, etc.) and builds a set of features. Features could be raw information, like location, or derivatives, like velocity (from location in time). Features may or may not be context depen- dent. During feature analysis, the system transforms this raw information and recognizes the underlying behavior. In most cases, feature extraction is a distributed problem, since device networks (e.g., video cameras) can provide information on the scene from diferent viewpoints. In our particular case, the domain is inherently distributed because we have a four- camera network providing information on a complex scene in which single viewpoints are frequently occluded. Te research reported in this paper is part of a larger project, aim- ing to discover high-level information (recognition of group activities) by extracting the features of the scene using a dis- tributed network of devices. We propose a new representation of the selected features that could be used to learn and predict group behavior. Although a scene may, in some domains, have a great many features and feature types may vary, we believe that the identifcation of the group elements is likely to be very meaningful in most potential application domains. Such problems could have a wide range of possibilities. For this reason, static context information is essential for our pur- pose. For instance, we need to know how many groups there are, how many elements the groups contain, and whether the number of groups or of group members is static or dynamic. All this static context information depends on the specifc domain. Also, some domains could have a dynamic context that could be incorporated to our representation to improve the accuracy of the classifcation. In the basketball domain,