Comparative study of orthogonal moments for human postures recognition

Human posture recognition has recently attracted a significant attention from the computer vision community. However, as in any pattern recognition problem, the features extracted from human posture images must be relevant; otherwise, the overall performance of the recognition system may be affected. Among the large number of existing features, the orthogonal moments have been successfully used in many image analysis and pattern recognition applications. However, to the best of our knowledge, their performances in human posture recognition have not been yet examined. Thus, the objective in this paper is to evaluate and compare the performances of various types of orthogonal moments, namely, Zernike, pseudo-Zernike, orthogonal Fourier-Mellin, Gegenbauer, exact Legendre, Chebyshev, Krawtchouk and Hahn moments for human postures recognition problem. The performance evaluations of these moments, as well as a comparison between them, are performed on three public datasets, namely Zhao & Chen dataset, URFD dataset and SDUFall dataset. The obtained results showed that, using moments up to order 8 on Zhao & Chen dataset, and up to order 6 on URFD and SDUFall datasets, Krawtchouk moments and Hahn moments outperform all the other moments, reaching an accuracy of about 98.5% to 99%. The robustness of the different orthogonal moments against noise and segmentation errors was also evaluated. The obtained results showed again the outperformance of Krawtchouk and Hahn moments compared to the other moments, with a performance drop of less than 1.46% in the presence of high noise level, and less than 6.2% in the presence of severe segmentation errors.