Improving spatio-temporal feature extraction techniques and their applications in action classification

Abstract

Space-time feature extraction is a recent and popular method used for action recognition. This paper presents a new algorithm to improve the robustness of spatio-temporal feature extraction techniques against the illumination and scale variations. Most of the interest point detectors are sensitive to illumination variations that may cause serious problems in action recognition by finding wrong keypoints. A method is proposed to make the 3-D Harris corner detector robust to illumination changes. Illumination invariance is achieved by applying a contrast stretching function to the video to find the proper intensity level for each pixel. A non-uniform binning method is also proposed to make the 3-D extension of the well-known SIFT descriptor more reliable and robust to scale changes by forming orientation histograms which concentrate on the regions near the interest points. Bag of features technique is used for classifying actions provided by the KTH dataset and the results demonstrate that our proposed method outperforms the original 3-D corner detector and SIFT descriptor.