A Combination of Moment Descriptors, Fourier Transform and Matching Measures for Action Recognition Based on Shape

Abstract

This paper presents an approach for human action recognition based on shape analysis. The purpose of the approach is to classify simple actions by applying shape descriptors to sequences of binary silhouettes. The recognition process consists of several main stages: shape representation, action sequence representation and action sequence classification. Firstly, each shape is represented using a selected shape descriptor. Secondly, shape descriptors of each sequence are matched, matching values are put into a vector and transformed into final action representation—we employ Fourier transform-based methods to obtain action representations equal in size. A classification into eight classes is performed using leave-one-out cross-validation and template matching approaches. We present results of the experiments on classification accuracy using moment-based shape descriptors (Zernike Moments, Moment Invariants and Contour Sequence Moments) and three matching measures (Euclidean distance, correlation coefficient and C1 correlation). Different combinations of the above-mentioned algorithms are examined in order to indicate the most effective one. The experiments show that satisfactory results are obtained when low-order Zernike Moments are used for shape representation and absolute values of Fourier transform are applied to represent action sequences. Moreover, the selection of matching technique strongly influences final classification results.