IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 2, NO. 3, SEPTEMBER 2007 623 Correspondence Gait Recognition Using Compact Feature Extraction Transforms and Depth Information Dimosthenis Ioannidis, Dimitrios Tzovaras, Ioannis G. Damousis, Savvas Argyropoulos, and Konstantinos Moustakas Abstract—This paper proposes an innovative gait identification and au- thentication method based on the use of novel 2-D and 3-D features. Depth- related data are assigned to the binary image silhouette sequences using two new transforms: the 3-D radial silhouette distribution transform and the 3-D geodesic silhouette distribution transform. Furthermore, the use of a genetic algorithm is presented for fusing information from different fea- ture extractors. Specifically, three new feature extraction techniques are proposed: the two of them are based on the generalized radon transform, namely the radial integration transform and the circular integration trans- form, and the third is based on the weighted Krawtchouk moments. Ex- tensive experiments carried out on USF “Gait Challenge” and proprietary HUMABIO gait database demonstrate the validity of the proposed scheme. Index Terms—Gait authentication, generalized radon transforms, genetic fusion, 3-D surface silhouette distribution. I. INTRODUCTION Gait analysis has recently received growing interest within the com- puter vision community. Human movement analysis emerged a few decades ago mainly for medical analysis purposes [1], [2]. The latest research activities in multimodal biometrics evaluate the use of gait as a promising biometric modality. From a surveillance perspective, gait is an interesting modality because it can be acquired from a distance inconspicuously. A. Current Approaches in Gait Recognition Present work on automatic gait recognition has focused on the development of methods for extracting features from the input gait sequences. Gait analysis can be divided mainly into two techniques: model based and feature based (model free). Model-based approaches [3]–[7] study static and dynamic body parameters of the human locomotion. In [3], a multiview gait recognition method was presented using static activity-specific parameters, which are acquired from automatic segmentation of the body silhouette into regions. In [5], a gait recognition method has been proposed based on a statistical shape analysis. Moreover, in [7], a method was presented that extracts the gait signature from the evidence-gathering process. Experimental analysis in a dataset of ten subjects exhibited encouraging results. Conclusively, model-based approaches [3]–[7] create models of the human body from the input gait sequences. Previous work on these approaches shows that they are view and scale invariant. However, experimental evaluation in larger publicly available databases is Manuscript received October 31, 2006; revised May 28, 2007. This work was supported by the EU Cofunded Projects HUMABIO and STREP, 026990. The associate editor coordinating the review of this manuscript and approving it for publication was Prof. Rama Chellappa. The authors are with the Informatics and Telematics Institute, Thermi– Thessaloniki 57001, Greece (e-mail: djoannid@iti.gr; tzovaras@iti.gr; damousi@iti.gr; savvas@iti.gr; moustak@iti.gr). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TIFS.2007.902040 needed in order to compare their performance to that of feature-based methods. On the contrary, feature-based techniques used for gait recognition do not rely on the assumption of any specific model of the human body for gait analysis. Initially, the binary map of the moving person is esti- mated and a feature vector is extracted from the silhouette sequences. In [8], the extraction of features was performed on whole silhouettes; in [10], width vectors were used; in [11], Fourier descriptors were in- troduced; and, finally, in [12], angular transform was applied in silhou- ette sequences. All of the aforementioned techniques employ the use of gait in a temporal manner. The final stage in the feature-based ap- proaches is the selection of the matching method to be used for finding the similarity between two input gait sequences. Proposed methods for matching are based on simple temporal correlation; full volumetric correlation on partitioned subsequent silhouette frames [8], [9]; linear time normalization [13]; and dynamic time warping [14], [15]. In most cases, Euclidean distance was used as a metric for distance calculation, but there are also reports on using procrustes distance [5] and sym- metric group distances [16]. B. Motivation—The Proposed Approach This paper proposes a novel gait identification and authentication method based on the use of novel 2-D and 3-D features of the image silhouette sequence. The proposed algorithm is tested and evaluated in two datasets and was compared to the state-of-the-art methods in gait analysis and recognition. Furthermore, the use of a genetic algorithm is proposed for fusing information from different feature extractors. Specifically, three new feature extraction techniques are proposed: two of them are based on the generalized radon transform, namely the ra- dial integration transform (RIT) and the circular integration transform (CIT), which have been proven to offer a full analytical representation of the silhouette image using only a few descriptors, and the third is based on the weighted Krawtchouk moments that are well known for their compactness and discriminating power. The use of moments for shape recognition has recently received great attention [4], [16], [24]. Lee and Grimson computed a set of image features based on moments [4]. Shutler [24] proposed the use of the Zernike velocity moments for describing the motion throughout an image sequence. Zernike mo- ments are based on a set of continuous orthogonal moment functions, such as Legendre moments. Experimental results on multiple datasets exhibited improvements in the recognition performance and illustrated their benefits over Cartesian velocity moments. One common problem with these moments is the discretization error, which increases as the order of the moment raises and, thus, limits the accuracy of the com- puted moments [25]. However, motivated by the successful use of these moments for gait recognition, this paper introduces the use of a set of discrete orthogonal moments, which do not involve any numerical ap- proximation and are based on the weighted Krawtchouk polynomials [25], [26]. As a result, the error in the computed Krawtchouk moments is nonexistent and a reliable reconstruction of the original image can be achieved using relatively low-order moments. By using these weighted Krawtchouk moments, the recognition performance on the Gait-chal- lenge database [9] will be seen to improve over the methods in [8], [13], [14], [32], and [33]. This paper also introduces the use of depth data, captured by a stereo camera for gait signal analysis. Depth-related data are assigned to the binary image silhouette sequences using two new transforms: the 3-D 1556-6013/$25.00 © 2007 IEEE