D.-Y. Yeung et al. (Eds.): SSPR&SPR 2006, LNCS 4109, pp. 596 – 603, 2006. © Springer-Verlag Berlin Heidelberg 2006 HMM-Based Gait Recognition with Human Profiles Heung-Il Suk and Bong-Kee Sin Computer Engineering, Pukyong National University {daedalos, bkshin}@pknu.ac.kr Abstract. Recently human gait has been considered as a useful biometric sup- porting high performance human identification systems. We propose a view- based pedestrian identification method using the dynamic silhouettes of a human body modeled with the hidden Markov model (HMM). Two types of gait mod- els have been developed both with a cyclic architecture: one is a discrete HMM method using a self-organizing map-based VQ codebook and the other is a con- tinuous HMM method using feature vectors transformed into a PCA space. Ex- perimental results showed a consistent performance trend over a range of model’s parameters and the recognition rate up to 88.1%. Compared with other methods, the proposed models and techniques are believed to have a sufficient potential for a successful application to gait recognition. 1 Introduction Recognizing people by their gait, the style of walking of an individual, can be per- formed without asking them to take any specific actions and even without making them be aware whether they are being watched or not. From Johansson’s studies in psychophysics with moving light displays (MLD) attached to body parts, it appeared that humans have the capability of recognizing their acquaintance only through their gait [1]. Recently human motion analysis has been receiving increasing attention from computer vision researchers and it is well explained in the review papers by J. K. Aggarwal et al. [2], C. Cedras et al. [3], and D. M. Gavrila et al. [4]. According to these papers two distinct methods for human motion analysis are distinguished: ‘model-based method’ using a priori shape models, and the other, called ‘appearance- based’ or ‘view-based’, without using explicit shape models. Both methods take a common sequential process of (1) feature extraction, (2) feature correspondence, and (3) high-level processing. The difference between them lies in the way of processing feature correspondence. ‘Model-based’ methods compare the input features taken from an input image with the parameters of the 2D or 3D body models prepared in advance, and make feature correspondence automatically. On the contrary, the ‘ap- pearance-based’ methods carry out the feature correspondence by varying the values of position, velocity, shape, color, and so on from consecutive frames. A brief review is in order. In the works of A. Kale et al. [5, 6], they used the width of the pedestrian’s silhouette of the binarized images as the feature vector and developed