Feature selection for activity recognition in multi-robot domains

Abstract

In multi-robot settings, activity recognition allows a robot to respond intelligently to the other robots in its environment. Conditional random fields are temporal models that are well suited for activity recognition because they can robustly incorporate rich, non-independent features computed from sensory data. In this work, we explore feature selection in conditional random fields for activity recognition to choose which features should be included in the final model. We compare two feature selection methods, grafting, a greedy forwardselection strategy, and ℓ1 regularization, which simultaneously smoothes the model and selects a subset of the features. We use robot data recorded during four games of the Small Size League of the RoboCup’07 robot soccer world championship to empirically compare the performance of the two feature selection algorithms in terms of accuracy of the final model, the number of features selected in the final model, and the time required to train the final model.