Human action recognition using action bank and RBFNN trained by L-GEM

Visual surveillance is widely used in monitoring, entertainment and public security in recent years. This arouses the growing demand of automatic analysis system to deal with large amount of data produced by video cameras. Human action recognition is one of the most popular topics in video analysis. However, human activities are extremely complex and the dimensions of features extracted from a video are very large. Hence, the construction of a highly accurate and fast classifier becomes one of the major challenging tasks in human action recognition researches. In this paper, we proposed an action recognition approach using a Radial Basis Function Neural Network (RBFNN) trained by the Localized Generalization Error Model (L-GEM). Representative feature vectors are extracted from videos by the Action Bank and then used as the inputs of the RBFNN. The reduction of uncertainty process is then applied to reduced noise from different classes. In our experiments, the proposed method outperforms SVM for human action recognition.