Vision-based gait recognition focuses on recognizing an individual with various features extracted from a video-sequence of the person walking. In comparison with other biometric features such as fingerprints, vision-based gait recognition has the advantage of being unobtrusive. For example, iris recognition systems have an obtrusive interface because an individual must exhibit specific behavior (e.g., gazing in a specific direction or remaining still during a recognition latency time) to be identified. However, vision-based gait recognition systems require no individual contact other than walking. So far, the gait is probably the only perceivable biometric feature from a great distance and at a low resolution in comparison with face recognition systems. Nixon et al. presented an extensive survey of the various vision-based identification methods based on human gait [6]. Most current vision-based approaches are based on analysis of silhouettes in the sequences of images of walking human subjects. They may be explicitly classified into two main categories: (i) model-based analysis and (ii) model-free analysis.

The model-based analysis aims to explicitly model the human body or motion, and they usually perform model matching in each frame of a walking sequence so that the parameters such as trajectories are measured on the model. BenAbdelkader et al. estimated the cadence and stride length from low-resolution video based solely on the periodicity of the walking person and a calibrated camera [7]. Bobick and Johnson extracted relative body parameters from action of waking to describe the subject’s body and stride [8]. Tanawongsuwan et al. deployed joint angle trajectories extracted from markers placed on joint positions in the legs and on the thorax [9]. Yam et al. explored the intimate relationship between walking and running that was expressed as a mapping based on the idea of phase modulation [10]. Zhang et al. used the change in orientation of human limbs [11].

The model-free analysis establishes a correspondence between successive frames based on the prediction or estimation of features related to position, velocity, shape, texture, and color. Kale et al. considered the width of the outer contour of a binarized silhouette [2] while Sundaresan et al. considered the entire binary silhouette itself [12]. Vega and Sarkar used the change in the relational statistics among the detected image features, removing the need for object models, perfect segmentation, or part-level tracking [13]. Liu and Sarkar developed an average silhouette as the simplest recognition feature [14]. Collins et al. deployed key frame analysis for sequence matching with innate viewpoint dependence [15]. Lee and Grimson deployed ellipsoidal fits to human silhouettes [16]. The feature vector is composed of parameters of moment features in image regions derived from silhouettes. Bhanu et al. considered kinematic and stationary features by estimating 3D walking parameters by fitting a 2D kinematic model to 2D silhouettes [17]. Similarly, Han et al. used gait energy image formed by averaging silhouettes [18].

The main difference between the floor- and vision-based approaches lies in the biometric sensor used in the vision-based approach: a camera. This unobtrusive and transparent interface of vision-based gait recognition provides a strong advantage over other contact-type biometric recognition approaches such as iris recognition. Unfortunately, however, a camera is not always an optimal approach, as it is sensitive to environmental factors such as shadow and light intensity. Moreover, camera surveillance can compromise an individual’s privacy. Such an intrusion in privacy-sensitive environments such as one’s home has been the target of criticism in vision research. In the floor-based approach, by contrast, a floor is used as a biometric sensor that gathers the various features of one’s gait. Therefore, the floor-based approach not only preserves individual privacy but also withstands the effects of shadow and light. Like the vision-based approach, floor-based gait recognition begins from the notion that gait is observable, distinguishable, and idiosyncratic. Because a floor is the biometric sensor, measurable information is obtained from only the soles of an individual’s feet. Thus, the features for recognition are somewhat different from those of the vision-based approach.

Stride length and cadence are attributed to physical makeup such as height, body mass, and the lengths of limbs. Actually, many vision-based studies have examined how stride length and cadence in human walking can be employed for gait recognition [7,19,20]. Demonstrating the effectiveness of stride length and cadence in discriminating the gait of individuals, these studies contended that stride length and cadence could be used as biometric identifiers. They estimated cadence and stride length from a video-sequence by using the periodicity of the walking person and calibrated cameras. This style of human recognition provided the motivation for the research into the feasibility of a floor-based identification system in which walking features such as stride length and cadence could be employed. The first use of walking features as a biometric identifier in floor-based gait recognition systems was UbiFloorI developed in [21]. Strictly speaking, the feature they used was not stride length but center positions of footprints over multiple footsteps. However, considering the longitudinal and transversal differences between consecutive center positions of footsteps represent the variation of stride length and dynamic range respectively, UbiFloorI might be the first floor-based gait recognition system exploiting walking features including stride length and dynamic range. UbiFloorI consists of 144 inexpensive ON/OFF switch sensors (14 cm by 2.5 cm) fitted onto cushioned carpet and a data acquisition (DAQ) board for measuring and transmitting the data from the sensors. They extracted walking patterns, including stride length, dynamic range, and foot angle from a dataset of switch sensors and use a neural network to identify unknown walking samples. The experimental results showed about 90% recognition accuracy with 10 subjects. UbiFloorI provides users with a transparent and user-friendly interface because the only task users need to do is to walk naturally. However, time information such as stance and swing time (or cadence) was unavailable because of the unique mechanical characteristics of the switch sensors and the low resolution of the floor. Middleton et al. developed a prototype floor sensor (0.5 m by 3 m) composed of 1536 force sensitive resistors (FSR) and 3 PIC 16F84A microprocessors [22]. They collected dataset from 15 individuals and extracted walking pattern such as stride length and cadence and stepping pattern such as time on toe to time on heel ratio. Based on the features, they achieved an 80% recognition rate. More recently, Qian et al. developed a high-resolution pressure sensing floor to obtain 1D pressure profile and 2D position trajectories of the centers of pressure (COP) of both feet, forming a 3D COP trajectories over a footstep. Based on the 3D COP trajectories, they extracted various features such as stride length and mean pressure, and showed an average recognition rate of 92.3% with walking dataset collected from 11 subjects and Fisher linear discriminant classifier [23,24].

When a body is in contact with the ground, the downward force due to gravity is reflected back onto the body as a reaction. This reflected force is referred to as ground reaction force (GRF). Several attempts to use GRF for gait recognition can be seen in [25–27]. However, the focus of these efforts has been in diagnose and therapy rather than recognition and identification. The first attempt to deploy GRF for human recognition was the active floor [28], which consists of a 4 × 4 array of load cells. Each load cell supports the corners of four adjacent floor tiles. Addlesee extracted features from the vertical ground reaction force traces derived from an individual’s footsteps and used a hidden Markov model to recognize differences in their stepping patterns. Experimental results showed that the active floor correctly identified the footsteps of 15 subjects with 91% accuracy. Orr and Abowd’s work with the smart floor resembles that with the active floor [29]. They extracted 10 features from the vertical ground reaction force profiles of individuals and used the nearest neighbor search in 10-dimensional feature space to identify unknown footsteps. The method correctly identified the footsteps of the 15 subjects with 93% accuracy. Suutala et al. demonstrated a floor composed of simple binary switch sensors to detect footsteps and extract walking characteristics, and showed 84% total recognition rate with walking profiles collected from 9 subjects and a Bayesian approach using a Gaussian Process classifier [30].

Another method of identifying a person through gait recognition is using a static footprint, which is a footprint left after a person passes. Kennedy showed the possibility of identifying individuals through the pressure areas on the soles of their feet [31]. Nakajima measured the pressure distribution of a footprint with a pressure-sensing mat [32]. They showed an 85% recognition rate with 10 subjects. In this study, they assumed that the footprint-based identification and verification system would not work for the security application, but it could work for personal recognition in a small group such as that found in a home environment. Unlike the static footprint, a dynamic footprint can be defined differently depending on its biometric sensors. An example of a dynamic footprint is the GRF profile mentioned above. Jung defined a dynamic footprint as a COP (center of pressure) trajectory while stepping heel-strike to toe-off [33]. They measured a quantized COP trajectory from a mat-type pressure sensor and used a hidden Markov model to create probability models for each foot of a user. Their experimental results showed an 80% recognition rate with 8 subjects. In this paper, we present a new dynamic footprint, an array of transitional footprints over the floor we developed for user identification.

The most recent approach in gait recognition is the portable sensor-based approach, in which accelerometers are normally used as biometric sensors. Like the floor-based approach, the accelerometer-based approach has several advantages. One is that it has an unobtrusive interface, and it does not compromise a user’s privacy. On the other hand, it requires users to carry or attach an accelerometer and a motion-recording device on their bodies for identification. In the sensor-based method, accelerometers attached to several parts of the body such as the hip, the lower leg, or the waist, measure acceleration signal characteristics while the user is walking. Features are extracted from the output of the device accelerations in vertical, forward-backward, or sideways directions. Examples of accelerometer-based gait recognition systems include [34–36].

The software modules we develop for extracting a user’s walking pattern from data sets fall into two categories: (i) left footprint extraction and (ii) walking feature extraction. The left footprint extraction software is used to search for all footprints in the datasets received while one is walking over UbiFloorII, as shown in Figure 6.

We create an 8 × 4 footprint model that covers all probable footprints and choose three features as follows:

The X index of the backmost sensor in a footprint

Figure 7 displays a footprint model extracted from the footprint in STEP 1 of Figure 6. As shown in Figure 7, the backmost sensor in a footprint becomes the seed sensor, from which the other features can be extracted.

We extract spatiotemporal walking features using the feature values obtained from left footprint extraction as inputs of the walking feature extracting software. We adopt seven walking features as follows:

FX and FY : the physical X and Y coordinates of the backmost sensor in a footprint

FX and FY represent the physical X and Y coordinates of the seed sensor in a footprint with the bottom-left corner of UbiFloorII as the origin. The coordinates com_FX and com_FY represent the center of the footprint based on the footprint model, as shown in Figure 8. Practically, com_FX and com_FY comprehend the user’s stride length, dynamic range, and foot angle. fStart and fEnd imply the user’s stance and swing time in walking. Finally, to create input vectors to the neural network, we need to generate the sequences of each walking feature in terms of the footsteps, such as [com_FX1, com_FX2, com_FX3, . . . ].

To extract the stepping pattern in each footprint, we need to analyze the temporal variation of each footstep. Transitional footprint extracting software is used to analyze the variation of transitional footprints from heel-strike to toe-off in each footstep. Based on the left footprint obtained from the walking pattern extraction procedure, we obtain an array of transitional footprints in terms of the occurrence times of the events (i.e., changes in the states of the sensors). The array of transitional footprints extracted from the STEP 1 footprint in Figure 6 is shown in Figure 9, where the digit values denote the occurrence times (seconds) of the events. Finally, to make a standard of the stepping pattern, we extract the array of sampled transitional footprints at a uniform sampling time from the array of original transitional footprints. Figure 10 shows the array of sampled transitional footprints extracted from Figure 9 with a sampling time of 0.04 (s).

For user identification, we adopted two stepping features as follows:

The left footprint

To create input vectors to our neural network, we need to convert the left footprint and the array of sampled transitional footprints to the vectors. In the case of the left footprint, each gray square in Figure 7 is represented by a “−1”, and each dark square is represented by a “1”. Then, to create the input vectors, we scan the 8 × 4 left footprint of one column (i.e., four sensors) at a time. For example, in Figure 7, the input vector to the neural network looks like [−1 −1 −1 1 −1 −1 −1 1 . . . ]. In the case of the sampled transitional footprints, the process is similar to that in the case of the left footprint, except that the number of rows is 8 multiplied by the number of elements in the array (e.g., 18 in Figure 10).

We use multilayer perceptron networks to identify individuals based on the extracted walking and stepping features [39]. Inspired by biological neurons, neural networks (along with their components) consist of simple neurons, and the connections among them determine the function of the network. The network is trained to perform special tasks such as modifying the weights and biases of a network by applying the learning rule. The learning rule used in this work is the back-propagation algorithm. If well-trained with the back-propagation learning algorithms, the multilayer perceptron can correctly classify the samples even if they are unknown during network training.

The neural network that we deploy for identifying users by their walking patterns has the same structure as the neural network for identifying users by their stepping patterns, and both neural networks consist of three layers, as shown in Figure 11: (i) an input layer N₁ with P₁ neurons, (ii) a hidden layer N₂ with P₂ neurons, and (iii) an output layer N₃ with P₃ neurons. In the neural network for identifying an individual’s walking pattern, the number of neurons in the input layer is equal to the number of walking features F_N times the number of walking steps S_N. Likewise, in the neural network for identifying an individual’s stepping pattern, if the number of the elements in the step features is F_N and the number of footsteps used is S_N, then the total number of input neurons, P₁ is equal to F_N × S_N. The number of neurons P₂ in the hidden layer N₂ is selected by experiments and the number of neurons P₃ in the output layer N₃ is equal to the number of users M. Because the neural network is trained with supervised learning, the learning rule is provided with a set of examples (i.e., the training set) of proper network behavior. We have used sigmoid activation functions trained with backpropagation.

The training procedure for our network is as follows. In the neural network for walking pattern-based identification, we need to perform preprocessing; that is, we must normalize the input values to the network. In contrast, in the neural network for stepping pattern-based identification, the preprocessing procedure is not needed in our neural network because all elements of the input vectors are digital values. Thus, the input vectors converted from the left footprints or the arrays of sampled transitional footprints will be direct inputs to the neurons of the input layer in our network. The learning rule is used to adjust the weights and biases of the network in order to move the network outputs closer to the targets. Given a test sample for the well-trained network, we can choose the index number of the output node with maximum output values as the user’s identification number. (1) User _ Number = max ( O 1 , O 2 , … , O M )where O_i denotes the output values of the i-th node, and M is the number of users (i.e., 10).

Our experiments consisted of capturing walking samples from 10 experimental participants volunteered from the academic community. Data was collected from 10 males. Participants ranged in age from 27 to 35 (see Table 1). For each subject, we gathered 50 walking samples. We collected a total of 500 walking samples from each participant walking on the UbiFloorII system. Depending on the user’s stride length, it takes from five or six footsteps to cross UbiFloorII. Therefore, we consider only the first five steps (STEP_1 – STEP_5). All the subjects were asked to walk as normally as possible and given the option to listen to soothing music. Moreover, they were asked to walk with bare feet, reflecting the daily custom followed in the Korean homes. In order to mitigate any variation in their gaits, we imposed all of these conditions on the participants while collecting the walking samples.

To verify the dominant walking features, we use five feature sets as inputs to the network. Table 2 shows the combinations of these features. In Case 1, coordinates FX and FY are the inputs to the network. This case is used as the standard for evaluating the results with those of the other feature sets. The other feature sets comprise the combinations of our walking features.

To verify the dominant footsteps in stepping features, we use four feature sets as inputs to the network. Table 3 shows the combinations of the features. In Cases 1 and 2, the odd and even footsteps of the first four footsteps, respectively, are the inputs to the network. In Case 3, the first two footsteps are the inputs to the network, and in the last case, all four footsteps are inputs. In the experiments using the array of sampled transitional footprints, various input arrays can be extracted from the left footprint in terms of the sampling times. Thus, we use five sampling times, as shown in Table 4, in which the elapsed times of all footsteps is assumed to be less than 0.8 s.

We first demonstrate how the number of hidden nodes influences the performance of the neural network. In an effort to decide the optimal number of hidden nodes, we run an experiment in which we increase the number of hidden nodes while keeping other parameters fixed and observe the resulting recognition accuracy. The results of one experiment with com_FX, com_FY, fStart, and fEnd features are shown in Figure 12. The left side of Figure 12 shows that about 30 hidden nodes are sufficient to obtain about a 95% recognition rate. We also conduct experiments to decide the epoch and the goal. The right side of Figure 12 shows that after 800 epochs, the mean square error is smaller than 10⁻³, so this value is set to the goal.

We present the results of the test with our feature sets and a comparative analysis in Table 5. In this experiment, the recognition accuracies are obtained by averaging 10 simulation results changing the seed value that determines the initial values of weights and biases of the network. First, we can note that the compensation procedure for com_FX and com_FY results in about a 10% improvement in recognition accuracy. Considering Cases 2 and 3, Case 3 is worse than Case 2 (i.e., without nSensor) because nSensor information has already influenced the compensation procedure for com_FX and com_FY. We are able to achieve about a 96% recognition accuracy when com_FX, com_FY, fStart, and fEnd features are used. Therefore, we conclude that the stance and swing time is also a dominant feature for user identification.

In [21], they chose the physical X and Y coordinates of the center of the footprint, which produced the stride length and the dynamic range for a given individual, as the walking pattern, illustrated in Figure 1(a). Even though they could not use stance and swing time information due to the low density of UbiFloorI, the experimental results exhibited a recognition accuracy of about 90%. For UbiFloorII, we used not only the features that we employed for UbiFloorI but also stance and swing time information, illustrated in Figure 1(b). To observe the potential of the new time information for improving recognition accuracy, we first performed the experiments without stance and swing time, which showed a recognition accuracy of approximately 90%, which is almost the same as that of UbiFloorI. However, we observed that the recognition accuracy could be improved to as much as 96% with stance and swing time. Here, we find it very important to note that except for stance and swing time, recognition accuracy with the previous walking pattern (the physical X and Y coordinates of the center of the footprint) might not improve any more with regard to the resolution of the floor.

As we did when identifying users by their walking patterns, we first demonstrate how the number of hidden nodes influences the performance of the neural network. In general, the greater the number of hidden nodes, the longer the computation time. The results in this experiment with the arrays of sampled transitional footprints (Case 4, sampling time = 0.04 s) are shown in Figure 13, the left side of which shows that about 30 hidden nodes are sufficient for about a 90% recognition rate. We also conducted experiments to decide the epoch and the goal. The right side of Figure 13 shows that after 115 epochs, the mean square error (MSE) is smaller than 10⁻⁵, so this value is set to the goal. In all these experiments, we train the network using Powell-Beale and scaled conjugate gradient algorithms.

We present the results of the tests with our feature sets and a comparative analysis in Table 6. The first column shows the recognition accuracies with the left footprints, and the other columns show the recognition accuracies with the arrays of sampled transitional footprints in terms of various sampling times. In these experiments, the recognition accuracies are obtained by averaging the 10 simulation results changing the seed value. The experimental results with the left footprints show that the left footprint, which expresses the static shape of the sole of the user’s foot in UbiFloorII, is not distinctive enough to recognize individuals. In the experimental results with the arrays of sampled transitional footprints, most of them show a more than 80% recognition rate. We are able to achieve about 92% recognition accuracy with the arrays of the sampled transitional footprints (Case 4, sampling time = 0.04 s). Considering Cases 1, 2, and 3, two footsteps are not sufficient to achieve more than 90% recognition accuracy without regard to the order of the footsteps and sampling time. In addition, we can note that by shortening the sampling time for extracting the array of sampled transitional footprints from the array of original transitional footprints, the performance of recognition could be improved. Nevertheless, the small sampling time requires numerous inputs to the network, causing heavy computational load. Thus, we should consider limiting the sampling time for extracting the array of sampled transitional footprints.

In the UbiFloorII system, we chose static footprints and an array of sampled transitional footprints, or dynamic footprints, which represent the temporal variation of the shape of the bottom of the foot making contact with the floor from heel-strike to toe-off in each footstep, as the stepping pattern, illustrated in Figure 2. The experimental results from the static footprints, which express the static shape of the sole of the user’s foot in UbiFloorII, showed that they are not distinctive enough for human recognition, an expected finding based on the previous study [32]. The experimental results with the array of sampled transitional footprints exhibited a recognition accuracy of about 92%, but some variation obviously occurred in the recognition accuracy as a result of the sampling time. By shortening the sampling time taken to extract, we determined that the array of sampled transitional footprints and the performance of recognition could be improved. Nevertheless, a small sampling time requires numerous inputs to the network, causing heavy computational load. Thus, we would have to consider limiting the sampling time for extracting the array of sampled transitional footprints.

The problem mentioned previously is the numerous input nodes in the neural network, causing heavy computational load, for example as in the case of the best recognition result, (2) #     of input nodes = 32   ( footprint model ) * 0.8   ( maximum stepping time ) 0.04   ( sampling time ) * 4   ( #     of used footsteps ) = 2560

One solution to this problem is training in principal components analysis (PCA). PCA is a way of identifying patterns in data and expressing the data in such a way as to highlight their similarity and differences [40]. The purpose of PCA training is to obtain several principal components that represent the original stepping features from a high-dimensional measurement space to a low-dimensional eigenspace. PCA training has been employed on the vision-based gait recognition approaches [41–43]. In the floor-based gait recognition approaches, Suutala and Röning built a frequency domain presentation of GRF signal (camel-back curve) and deployed PCA to reduce high dimensionality of the amplitude spectrum presentations [44]. We adopted an analogous method to that presented in [42], reducing the original 2560-dimensional space to a vector with 248 coefficients. The accuracy of recognition with this PCA training technique was about 89%, which is reasonable considering the greatly reduced computational loads.

Table 7 summarizes mean and standard deviation for classification accuracy over the selected classification methods based on walking pattern we extracted from the walking samples. We used 10 times 10-fold cross-validation, i.e., 10 different 10-fold cross-validation experiments with the same learning method and dataset, averaging the 100 experimental results. Overall, recognition accuracy is highest for multilayer perceptron, which is consistent with the experimental result with MATLAB in the previous section. Support vector machine is the second most accurate algorithm, and k-nearest neighbor method shows good performance as well. However, rule-based learning algorithms, particularly decision table shows lower accuracy compared to other algorithms. After looking into the confusion matrix, we know that this is because some portion of the walking samples of each subject was misclassified as a particular user (User₀), i.e., False Positive rate of the subject is considerably higher than others.

Table 8 shows an aggregate confusion matrix for multilayer perceptron classifier based on 10 subjects’ walking pattern. Recognition accuracy for User₀ is lower than others because some portion of the walking samples of User₀ was misclassified as User₁, User₃, User₄, User₉. We can also know that User₂ and User₃ are confused with User₇ and User₆ respectively since the subjects may have walked in similar style while collecting walking samples. However, it should be noted that the confusion between the users might not have been occurred by their height since there is very little correlation between them as shown in Table 1.

Table 9 summarizes mean and standard deviation for classification accuracy over the selected classification methods based on stepping pattern we extracted from the walking samples. We used 10 times 10-fold cross-validation, i.e., 10 different 10-fold cross-validation experiments with the same learning method and dataset, averaging the 100 experimental results. Overall, recognition accuracy is highest for support vector machine. It should be noted that this result is not very surprising because the past work in [44] recognized 11 subjects’ different single-footstep feature presentation with support vector machine classifier, and its recognition accuracy was higher than multilayer perceptron neural network. Multilayer perceptron is the second most accurate algorithm, and its accuracy is consistent with the experimental result with MATLAB in the previous section. As in case of walking pattern-based user identification, rule-based learning algorithms show lower performance than others because False Positive rate of User₀ is considerably higher than others. It also should be noted that Bayes net shows weak performance compared to the discriminative model, which is consistent with past works on footstep-based recognition.

Table 10 shows an aggregate confusion matrix for support vector machine based on 10 subjects’ stepping pattern. We can know that User₅ is often misclassified as User₄ since the subjects may have foot-stepped in similar style while collecting walking samples. The confusion between users might not have been occurred by their foot size because there is very little correlation between them as shown in Table 1.

In the above section, we showed that both the walking and stepping patterns extracted from the human gait can be employed for the biometric identification of an individual. However, the recognition accuracy could be improved by combining two classifiers. In the last decade, a substantial amount of research has been devoted to multimodal biometric systems [47,48]. Such systems exhibit high performance when the biometric features stem from totally disparate sources such as faces, voices, and fingerprints. In the floor-based gait recognition method, Suutala and Röning presented a way of improving classification accuracy and the adaptability based on the conditional posterior probability outputs of classifiers, i.e., efforts to combine classifiers trained with different feature sets and to combine multiple footstep instances of a single person walking on the floor [44].

Disparate biometric identifiers are fused using three basic techniques [49]: (i) fusion at the feature extraction level, (ii) fusion at the matching score level, and (iii) fusion at the decision level. The dimension of the feature vector of the stepping pattern is significantly higher than that of the walking pattern. Thus, it is not reasonable to concatenate the two feature vectors into a single vector for fusing at the feature level. Instead, we have adopted fusion at the matching score level. Figure 14 presents a flow chart of two classifiers combined by fusion at the matching score level for gait recognition.

In our case, each neural network generates ±1 max-min score values, so normalization of the output score is not necessary. We can combine the output score values from each system into a new score value using Equation (3): (3) O G ( i ) = W W × O W ( i ) + W S × O S ( i )where O_G(i) denotes new output values of the i-th node, and W_W and W_S represent the weighting parameters for the walking and stepping patterns, respectively. The weighting values W_W and W_S were chosen so that the recognition accuracy could be maximized by combining the two classifiers (in our experiment W_W = 0.96, W_S = 0.92). O_W(i) and O_S(i) depict the output values from the neural networks of the walking and stepping patterns, respectively. Finally, we can choose the index number of the output node with the maximum output values as the user’s identification number (See Equation (1)). With this fusion technique, we could achieve a recognition accuracy of about 99%. From the experimental results, however, we cannot claim that our gait recognition technique will almost perfectly recognize all the participants in a domestic environment. Instead, we choose to focus on the potential of combining two classifiers for the robust performance of the identification system since the human walk, or gait, is likely to vary over time.