Gait Recognition Using Fractal Scale and Wavelet Moments Guoying Zhao 1,2 , Li Cui 3 , Hua Li 2 , Matti Pietikäinen 1 1. Machine Vision Group, Infotech Oulu and Department of Electrical and Information Engineering, P. O. Box 4500 FI-90014 University of Oulu, Finland E-mail: {gyzhao, mkp}@ee.oulu.fi 2. Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100080 E-mail: lihua@ict.ac.cn 3. School of Mathematics Science, Beijing Normal University, Beijing 100875 E-mail: licui@bnu.edu.cn Abstract Video-based gait recognition is a challenging problem in computer vision. In this paper, fractal scale wavelet analysis is applied to describe and automatically recognize gait. Fractal scale based on wavelet analysis represents the self-similarity of signals, and improves the flexibility of wavelet moments. Optimal wavelets based on generalized multi-resolution analysis are used to improve the recognition rate. Descriptors of fractal scale are translation, scale and rotation invariant. Moreover, a combination of fractal scale and wavelet moments improves the recognition rate. Experiments show that the proposed descriptor is efficient for gait recognition. 1.Introduction Gait recognition is used to signify the identification of individuals in image sequences ‘by the way they walk’. Studies in psychophysics reveal [2] that humans have the capability of recognizing people from even impoverished displays of gait, indicating the presence of identity information in the gait signature. There are many proposed methods contributing to the gait analysis, for example model-based [1,3-5], and appearance-based [6-10]. Translation and scaling of walking people occurs often in video sequences, and moments are an efficient tool to deal with them. The application of classical moments to two dimensional images was first shown in the early sixties by Hu [11]. Little and Boyd used moment based features to characterize optical flow for automatic gait recognition [12], thus linking adjacent images but not the complete sequence. Lee and Grimson computed a set of image features based on moments [3]. Liu et al. used the first and second moments of two binary silhouettes to determine an affine transformation that coarsely aligns them [13]. Shutler and Nixon developed new Zernike velocity moments to describe the motion throughout an image sequence to help recognize gait [14]. The original definition for those moments apply global functions, so they only extract global information from images, which is just appropriate for discriminating two signals with apparent difference. Moreover, the low order moments describe the global structure and the details are reflected in the high-order moments. However, computing high-order Zernike moments and geometric moments is very time-consuming. Thus, it is not easy to correctly classify similar image objects with subtle differences based on such global and low order moment invariants and to be able to deal with noise. To avoid these problems, Zhao et al. proposed an approach based on wavelet velocity moments for gait recognition [10]. In the expression of wavelet moments given in [15], only the wavelet function with explicit continuous expression can be used. This limits the applicability, and also the computational complexity of that method is a bit high due to the lack of Mallat algorithm. In this paper, a new fractal scale descriptor based on discrete wavelet analysis is proposed, which trains optimal wavelets with the idea of generalized multi-resolution analysis (GMRA), and aims to describe the self- similarity of gait appearance through a time varying sequence. Combined features of fractal scale and wavelet moments improve the results of gait recognition. The method is computationally simple and improves the flexibility of wavelet moments. The 18th International Conference on Pattern Recognition (ICPR'06) 0-7695-2521-0/06 $20.00 © 2006