Abstract
Recognizing human actions from video has attracted a lot of attention from computer vision research community due to its many prospective applications. Feature extraction is a key step for achieving an efficient human recognition system. Among other feature extraction methods, complex Zernike moment (ZM) considered as one of the most simple and successful approach. Commonly, Zernike moments are employed for feature extraction using only magnitude value of the complex ZM. In this paper, we study the performance of combining various complex ZM parts such as: magnitude, real, imaginary and phase information to improve the recognition accuracy. The proposed methodology consists of five stages: motion energy image (MEI), Zernike moments (ZM), whitening transform, bag-of-features and support vector machine (SVM) algorithm. The proposed method achieve comparable results with other state-of-the-art methods using the publicly available Weizmann action recognition dataset.