The techniques of classifying human age using gait generally enlist features like static and dynamic gait behaviors. For instance, Davis [9] used the features like the length of legs, the width of stride, and the frequency of strides to categorize the gait into children and adult age groups. Begg et al. [10] have categorized people as younger and aged by applying the data of foot clearance. Chuen et al. [26] showed the usage of parameters like head-to-body ratio and length of legs for the classification of children's age and adult age. Nabila et al. [27] have highlighted the spatiotemporal features, such as arm swing, hunched posture, and length of stride, to show the difference between younger and elderly people. Xu et al. [24] have proposed a method for handling uncertainty problems in various age groups. Li et al. [15] have analyzed the human growth process and divided the age into 5-year intervals, resulting in nine age groups. They have proposed an age group-dependent framework to handle the large estimation errors (when variation in age is large). Mansouri et al. [28] used a fusion of gait contours and silhouette descriptors to categorize gait images into young and elderly.

Contrary to the above-mentioned research works, few studies show the use of appearance-based methods for the classification of various age groups. For instance, Mannami et al. [16] utilized the frequency-domain characteristics and categorized the person’s age into 3 groups: children (below 15 years), adults (in the range of 15 to 65 years), and the aged (more than 65 years). Abirami et al. [29] have combined gender and age for age estimation. They used Hilbert-Schmidt Independence Criterion to maximize the correlation between gender and age. However, it can be observed that age group classification methods have some drawbacks related to age groups and insufficient experimental validation.

Aderinola et al. [30] have conducted an extensive survey of scientific literature on gait-based age estimation from 2001 to 2021. In this survey, they followed similar age groups as the study [16] for the discussion. They concentrated on vision-based and sensor-based approaches for gait-based feature extraction. They also showed the effect of different covariates on vision-based and sensor-based gait data. Furthermore, the findings of the study include the following: 1) Model-free gait descriptors, i.e., GEI, are more robust than model-based gait descriptors. 2) There is a research gap for age estimation under the viewing angle variations. Hence, this study adopts the GEI and considers multiple viewing angles for age estimation.

Estimating an age based on gait behavior is the latest advancement in gait-related studies, and therefore, not much work is done in this field. Recently, some research work has begun on age classification based on gait. The earliest study on age estimation using gait behavior is done by Lu and Tan [8]. They encoded age as binary and applied multi-label k-nearest neighbors for classification. Thereafter, Makihara et al. [11] used Gaussian process regression (GPR) and estimated the age of persons. They combined the face recognition method with gait behavior. Lu and Tan [6] used analytical methods to discover a low-dimensional feature for age estimation. Marin-Jimenez et al. [19] proposed a multi-task CNN model, which takes input as a fixed-length optical flow sequence and gives output as various biometric traits for age estimation, person identification, and gender recognition. Zhang et al. [20] used a deep ConvNet to capture the gait characteristic from GEI. In order to achieve better accuracy in age estimation, they have applied a multi-task learning framework. Sakata et al. [23] used a deep learning framework (DenseNet) for gait-based age classification. They used the world’s largest gait dataset having age variations up to ninety years. Their study shows superior results in the classification of age. Riaz et al. [31] have gathered the inertial data of gait from 86 different persons and performed the analysis for age estimation. During data gathering, they considered angular velocities and 6D accelerations of a person. They mounted inertial measurement units on the chest for recording the inertial data. Furthermore, they divided the recorded data (long sequences of inertial signals) into single steps and evaluated 50 spatio-spectral features from every step. Finally, they trained the features with MLP, SVM, and random forest classifiers. These trained data were used for both age estimation and person recognition.

After studying the above-mentioned research works, we have identified that convolution neural network (CNN) based methods are more used in recent works, and they have also given better results for gait-based human age estimation. After going through the experimental setup of the CNN-based study, it can be learned that the computation cost of the system becomes very high as it demands high processing power. The other limitation is that these methods are carried out by considering only one viewpoint for age estimation. Thus, to overcome these limitations, our study has considered implementing a lightweight system for gait-based age estimation under various viewpoints. This study also tries to authenticate the results by comparing them with other studies done for age estimation using gait behavior.

Fig. (1) elucidates the proposed technique. The detailed functioning of the proposed system is explored in the forthcoming subsections.

Fig. (1). Overview of proposed system.

We have obtained silhouettes from a person’s walking sequences. Silhouette is a black and white image having the outline of a person in the foreground. The appearance-based method includes processing silhouettes using gait. As a person’s clothes, color and texture influence can be avoided using the appearance-based method; it is superior to the model-based method for a person’s gait identification. Here, we have followed the most popular representation for silhouette-based gait, i.e., GEI [32, 33]. We have evaluated GEI from a sequence of silhouettes. In this experiment, we have fixed the GEI size as 88 x 128 pixels. GEI is very simple and, at the same time, yields highly effective gait features. GEI is estimated as follows:

(1)

where S_t(x,y) indicates the binary gait silhouette, t is the counting variable for the gait frame, x and y are coordinates of the 2D image, and N_G. is the number of gait frames extracted from one cycle of the periodic walk.

Fig. (2). Gait energy image.

Fig. (2) illustrates the formation of GEI from gait silhouettes. GEI comprises both physiological and spatiotemporal information about the human body during walking. It can be observed that GEI representation saves computation time and storage space for recognition. Let S_t (x,y) is formed by adding distortions e_t (x,y) to an actual gait picture frame g_t (x,y). Thus, S_t (x,y) = g_t (x,y) + e_t (x,y). Assuming that the distortions in the gait picture are evenly spread and not related to t, e_t (x,y) can be computed from the following equation:

In N_G. frames, let R picture frames have g_t (x,y) = 1 for the position (x,y). Thus Eq. (1) will be specified as:

(2)

Therefore, GEI noise can be represented as:

(3-7)

Thus, the mean of the noise in GEI toggles between -m and m with various values of R. If R toggles between 0 and N_G. at position (x,y), E {ē(x,y)} toggles between m and -m. Therefore, both σ²_ēt (x,y) and E {ē(x,y)} of the noise for GEI are suppressed in comparison with the solo silhouette image on the particular positions. This proves the significance of GEI over a solo silhouette for obtaining gait features.

A detailed comparison has been shown in a study to find the best-suited transform method for feature extraction in pattern recognition applications [34]. The outcome of the study highlights that the performance of DCT is superior to the other transform methods (Karhunen-Loeve Transform, Discrete Fourier Transform, Walsh-Hadamard Transform, and Haar Transform). The study [35] summarizes that DCT is better than DWT (Discrete Wavelet Transform) in terms of energy compaction, computational complexity, and performance time over the image data. Therefore, after GEI formation, we have applied DCT to it for feature extraction. Earlier, DCT was used for image compression [36, 37]. However, over the years, it has been observed that the pattern recognition community has shown more interest in DCT [38-41]. Basically, the DCT method is used to convert image data into frequency components [42]. It accumulates higher coefficient value components into the top left corner in the 2D matrix representation, whereas the lower coefficient values are kept in the bottom right of the 2D matrix. The following equation gives DCT matrix elements:

(8)

Where N denotes the block size, and g(m,n) represents the image in matrix g showing m,n^th matrix element.

The DCT coefficient at the (x,y) point in the DCT domain is denoted by DCT (x,y). The DC coefficient DCT(0,0) and the AC coefficients make up the DCT coefficients. The AC coefficients are used to calculate the focus value. Here, the primary goal is to recognize the optimal feature and extract it using DCT by reducing the dimension of GEI data.

Extraction of features using DCT contains 2 phases. The initial phase uses DCT on the complete GEI picture frame to get coefficients of DCT. The latter phase selects the coefficient with a high frequency, which is utilized for generating feature vectors as input to MLP.

Fig. (3) illustrates a pictorial representation of DCT applied to the GEI image. Fig. (3) consists of 3 blocks where the leftmost one is the GEI image, the rightmost top block represents DCT transformed picture, and the right side bottom block shows the categorization of coefficients of DCT. Generally, the coefficients of DCT are categorized into 3 sections: lower frequency components, moderate frequency components and higher frequency components. The lower frequency components are dependent on the light intensity of the surrounding, whereas other components contain meaningful data and can rebuild the picture again; thus, these components are acceptable for feature vector generation and can be used in gait identification.

During model training, for removing the outliers, normalization of feature vectors is carried out. Thus, a feature vector is transformed to have a mean of 0 and a variance of 1 in a normalized feature vector. Fig. (4) shows the flowchart of the tasks involved in MLP. For training the model, feature vectors that are normalized are given as input to the hidden blocks. The number of neurons in the hidden blocks is 700, 500, 200, 100, and 50, forming a completely linked dense node, respectively.

Fig. (3). Illustration of DCT on GEI with pictorial representation.

Fig. (4). Illustration of tasks involved in MLP.

The proposed MLP model performs training by making smaller batch sizes and calculates the gradients of these batches. The purpose of adding a batch normalization layer is to increase the performance as well as consistency of each hidden layer's smaller version. The internal covariate shift can be reduced by eliminating gradient descent oscillation. Using the batch normalization approach, the output of each layer's mini-batch is modified, and it can be put into the following layer. During this training, we have run 1,000 epochs. Furthermore, to enhance the working of the MLP model, we have chosen an optimizer as RMSprop having a learning rate of 0.002 to tune the hyper-parameters.

We have assessed the performance of age estimation using the gait feature. In this assessment, we have calculated a mean absolute error (MAE) between the actual input age (ground truth age) and system estimated age. The ground-truth label can be explained with the help of a probability distribution (discrete).

Let be the estimated discrete probability distribution for the t-th training data (1,2…..,N_g) where N_g is total training data, k as age and as the likelihood for k-th age label.

We evaluate the mean age from the distribution and measure the mean absolute error (MAE) as per the following equation:

(9-10)

Where, ŷ_t is expected age and y_t is actual age for t-th training data.

We have tried to show the authenticity of the proposed system by matching its execution with state-of-the-art methods. Here, MAE (in years) is used as the metric for comparison. We have compared the gait-based age estimation results with both the conventional methods and deep learning-based methods. The conventional methods used for comparison are SVR (Gaussian), SVR (linear), AGDMLR [15], OPMFA [6], OPLDA [6], and GPR with k nearest neighbors and results with k = 10; 100; 1000 [11]. Fig. (6) illustrates that the proposed system outperforms the other conventional methods for gait-based age estimation. The deep learning-based methods used for comparison are DenseNet [23], Multi-task [20], Multi-stage [44], GEINet [45], GEINet [24], GaitSet [24]. Fig. (7) shows that the performance of the proposed system is comparable with all the deep learning methods mentioned here. The experimental results signify that DCT can effectively decide different frequency components. DCT is inconsiderate of the variations in human appearances. Furthermore, the tuned MLP increases the performance of age estimation.

After tuning the MLP, MAE for age estimation under every viewing angle is evaluated. To the best of our knowledge, only Xu et al. [25] have shown angle-wise performance analysis of age estimation. Therefore, this study compares its result to their findings. Table 1 shows the comparative analysis of the performance of age estimation under viewing angle variations. It is observed that the proposed system attains the best MAE of 5.05 for age estimation under a viewing angle of 90°. It is obvious that in the side view of a person, the majority of the gait features are observed, which helps in age estimation. At 0° viewing angle, comparatively fewer features are observed; hence the MAE rises to 7.47. It can be observed from Table 1 that the results of the proposed system are superior to their study’s findings [25]. The study [25] used a deep learning approach to estimate the age of a person, whereas we used a lightweight approach. The results shown in Table 1 indicate that the model-free gait descriptors (GEI) have huge potential in gait recognition.

Fig. (6). Comparative analysis of performances of conventional ML methods with the proposed method.

Fig. (7). Comparative analysis of performances of deep learning methods with the proposed method.

It can also be seen that the proposed system shows outstanding performance for every viewpoint, especially at 90°, 270°, 255°, 75°, 30°, 45°, 60°, 240°, 225°, and 210°. Thus, the proposed system shows robustness to viewing angle variations.

To prove the authenticity of the proposed system, we have performed a statistical analysis of recent works done in the gait-based age estimation field. The outcomes of statistical analysis are represented in the tabular form in Table 2. We have considered the benchmarking work for both classification and regression-based age estimation. It was observed that conventional machine learning algorithms could perform better in smaller datasets. Most of the studies using conventional machine learning follow hybrid feature extraction methods. The hybrid features are a fusion of kinematic and biological features of a gait. Generally, the hybrid feature extraction techniques work on a handcrafted dataset, where the setup for recording the biological gait features is maintained. However, the size of the dataset is small. To verify the functioning of the system, a larger dataset is needed. The recent deep learning approaches have used vision-based inputs; hence, they used kinematic features to perform the age estimation. These methods have verified their results on larger datasets like OULP-Age, OULP, and OU-MVLP. It is observed that these methods have used a regression-based approach and attained a noteworthy performance. However, the proposed system has adopted a lightweight approach and attained a comparable performance to deep learning methods.