Recognizing realistic action using contextual feature group

Abstract

Although the spatial-temporal local features and the bag of visual words model (BoW) have achieved a great success and a wide adoption in action classification, there still remain some problems. First, the local features extracted are not stable enough, which may be aroused by the background action or camera shake. Second, using local features alone ignores the spatial-temporal relationships of these features, which may decrease the classification accuracy. Finally, the distance mainly used in the clustering algorithm of the BoW model did not take the semantic context into consideration. Based on these problems, we proposed a systematic framework for recognizing realistic actions, with considering the spatial-temporal relationship between the pruned local features and utilizing a new discriminate group distance to incorporate the semantic context information. The Support Vector Machine (SVM) with multiple kernels is employed to make use of both the local feature and feature group information. The proposed method is evaluated on KTH dataset and a relatively realistic dataset YouTube. Experimental results validate our approach and the recognition performance is promising.