Human biometric analysis has gotten much attention due to its widespread use in different research areas, such as security, surveillance, health, human identification, and classification. Human gait is one of the key human traits that can identify and classify humans based on their age, gender, and ethnicity. Different approaches have been proposed for the estimation of human age based on gait so far. However, challenges are there, for which an efficient, low-cost technique or algorithm is needed. In this paper, we propose a three-dimensional real-time gait-based age detection system using a machine learning approach. The proposed system consists of training and testing phases. The proposed training phase consists of gait features extraction using the Microsoft Kinect (MS Kinect) controller, dataset generation based on joints’ position, pre-processing of gait features, feature selection by calculating the Standard error and Standard deviation of the arithmetic mean and best model selection using R^{2} and adjusted R^{2} techniques. T-test and ANOVA techniques show that nine joints (right shoulder, right elbow, right hand, left knee, right knee, right ankle, left ankle, left, and right foot) are statistically significant at a 5% level of significance for age estimation. The proposed testing phase correctly predicts the age of a walking person using the results obtained from the training phase. The proposed approach is evaluated on the data that is experimentally recorded from the user in a real-time scenario. Fifty (50) volunteers of different ages participated in the experimental study. Using the limited features, the proposed method estimates the age with 98.0% accuracy on experimental images acquired in real-time via a classical general linear regression model.

Age is a significant human characteristic that can be used to evaluate a person’s identification, security, and health [

The structure of this paper is as follows: In Section 2, the literature review is covered, Section 3 presents the proposed system, and Experimental work is described in Part 4. Section 5 describes the comparison to the state of the art, and finally, Section 6 described the conclusion.

Age Calculation using gait is a modern research, therefore, limited investigations have been carried out so far. The first study on this area was published by Lu et al. [

The main contributions of the proposed research work as follows: (1) preparation of a new dataset for age estimation; (2) Extraction and selection of features using different statistical analyses such as T-test, ANOVA technique, R^{2} and adjusted R^{2} (adj-R^{2}) techniques; (3) Using a simple linear regression model for the complex problem of age estimation. The proposed model is computationally optimal with very low computational cost and highest accuracy; (4) the proposed method can be applied in a real-time scenario for the new user, with data extracted from MS Kinect; (5) the proposed system is compared with the state-of-the-art system using different parameters such as accuracy, technique, and dataset used.

The proposed system consists of two main parts: training and testing. The proposed training phase consists of different components, i.e., extraction of Skeletal information through Microsoft Kinect (MS Kinect), dataset generation, Pre-processing, feature selection method, best model selection, and finally age Estimation is carried out using the best-fit model (see

The proposed testing phase automatically estimates the age of a walking person in real-time using the results obtained from the training phase (See

We created our own dataset because there were no other datasets that contained both skeleton joint position and were large enough gender information. In this research work, 273 individuals, all 7–70 years old, participated in the experiment. The authors asked each volunteer to walk in a free pose in different directions inside the Kinect recognition space (see

To record the 3D joints position from volunteers, a low-cost Microsoft Kinect V1 camera is used. The specification of the Kinect camera used in terms of resolution is (640 × 480). The depth information between the objects and the camera is captured by this sensor in addition to the conventional RGB image of the depicted object. The MS Kinect consists of a 3D depth sensor, a RGB camera, a tilt motor, and a Microsoft array as shown in

The accuracy of MS Kinect was tested by recording multiple records for each of the subjects. The measurement was determined from different distances and the optimal distance was determined to be 6 feet. The height and angle of the MS Kinect were 3.4 feet and 0 degrees at 30 frames per second (fps), respectively. All volunteers were free to take any action or position at any time. Because the average of all frames was used to extract the feature, there was no time limit.

The authors used 20 joints to capture the functionality in the classification phase. The position of the joint P_{j} is defined as: P_{j} = x_{i} + yi + z_{i,} where P_{j }= 1, 2, 3…20 and i = 1, 2, 3…n. In this phase, the extracted data is pre-processed before feeding the machine learning tools. Pre-processing includes calculating the arithmetic mean of joints for each axis, and then calculating the combined arithmetic mean for each joint.

As the Kinect sensor provides 30 frames per second, approximately, and each joint position is represented in X, Y, and Z coordinates for each frame. Therefore, the arithmetic mean for joints is computed as

All body features were evaluated collectively and individually for each classification method throughout the classification phase. However, using all body features in the classification can result in a lower percentage of accuracy. As each joint moves in X, Y, and Z coordinates, therefore all the combinations of 60 features were tested. On the other hand, to improve classification accuracy, it is important to combine the features according to best model selection criteria R^{2} and adjusted R^{2}. The best model was identified based on the percentage of accuracy for most contributing joints. We checked each percentage of accuracy in the classification based on standard error. If the accuracy rate exceeds 90%, the individually classified feature will be selected and combined with the other selected features. Otherwise, the features were not added to the selected set of features. The standard error measures accuracy. Higher standard errors cause less accuracy. The standard error of the arithmetic mean for shoulder right-joint can be calculated as

The standard error and standard deviation can be represented in the same way using

The coefficient of determination (R^{2}) and (Adj – R^{2}) techniques are used to select the best fitted model for age estimation. To determine the best fitted model for which R^{2} is high i.e., if R^{2} = 90%, then 90% age is explained by joints and 10% due to other factors such as height, etc. Mathematically, R^{2} can be written as

Where y_{i} is the value of a walking person age, ^{2} is used to select a best fitted model. Mathematically, the adjusted R^{2} is defined as
^{2} and Adj-R^{2} lie between 0 and 1. When results of R^{2} and Adj-R^{2} are near to zero, tells us weak relationship and near to one, tells us strong relation between dependent (Joints) and independent variables (age).

The classification table for CLRM is used to find which human joints contribute the most to age detection. The most-contributed joints will be determined based on their accuracy. The most-contributed joints with an accuracy rate of more than 90% will be considered on the basis of standard error. Normally, the relationship among variables is direct, inverse and curving. To capture the relationship between variables, classical linear regression models (CLRM) are widely used. The linear relationship between one of the dependent variables, and one or more of the independent variables is known as regression. The CLRM is used when the dependent variable is continuous and follows normal distribution. In this study, the dependent variable is the age of a walking person and is continuous, which satisfies the important assumption of normality. Also, the Independent variables are the contributions or movements of 20 different human joints during gait. The results show that only nine joints out of twenty joints are the most heavily contributed joints in age estimation.

The CLRM is mathematically defined as
_{i} is the dependent variable, _{0} is intercept form, _{i} are the slop parameter, X_{i} s are independent variables, and

Estimating the model parameters of the classical linear model (CLM) is one of the key requirements. The least square is used because the least square estimators are more accurate and effective for estimating a model’s unknown parameters. We can write the CLRM in the form of matrices as

Taking _{0}, _{1}, _{2}…_{p} common
_{1} = Foot left and X_{2} = Foot right). Then

Also

Now the least squares normal equations

By solving

The hypothesis about regression coefficients is either significant or insignificant, i.e., _{i} are

To predict the age of a walking person on the basis of data obtained form 3D Kinect and regression, the CLRM is used. For calculating the age of a single subject, CLRM can be written as
_{0} ,β_{1} . . . β_{9} are the least square estimators and

An experimental study is conducted on a self-created dataset that contains gait features of 273 subjects. As we are interested in age factors, this dataset has age ranges from 7 to 70 years. There are several advantages of this dataset. This dataset is larger in size than the other self created dataset, providing enough subject detail for almost all the tests. Beside, a wide range of ages also gives an ideal dataset for testing the result. Furthermore, the dataset is consisted of 3D joint positions, while other contains 2D joint data. Therefore, the proposed dataset is ideal and verify the reliability of the proposed system. We carried out experimental analyses to assess the accuracy of different joints set for the age classification.

To ensure that the proposed system is feasible, we carried out an experiment via MS Kinect. We placed Kinect on a table of 3.0 feet height. Fifty (50) volunteer subjects participated in the experimental study. The path and direction to be followed during the experiment were initially demonstrated to the subjects. The task consisted of casual walking of a subject in front of the MS Kinect in a horizontal direction while following a predetermined path in the Kinect recognition space (see

As a classification method, we used CLRM because age is a continuous variable. For testing the validity of the CLRM, the ANOVA test is used, and all our data seem relevant and highly significant.

The results are presented in this section from the experimental evaluation. To test the validity of the proposed system, we performed different experiments using 50 walking subjects in different directions inside the Kinect recognition space. Before starting the experiment, the actual ages of the subjects were recorded. All of the subjects were assigned to one of five groups: A, B, C, D, or E (See

Group | Training | Testing | Age range | Subject calculated | Accuracy |
---|---|---|---|---|---|

A | 55 | 10 | 07–20 | 09 | 90% |

B | 55 | 10 | 21–30 | 10 | 100% |

C | 55 | 10 | 31–40 | 10 | 100% |

D | 55 | 10 | 41–50 | 10 | 100% |

E | 53 | 10 | 51–70 | 10 | 100% |

The T-test is used to determine the importance of a specific joints. In _{r}, _{r}, _{r}, _{l}, _{l}, _{l}, _{r}, _{r}, _{r} are the combined averages of right shoulder, right elbow, right hand, left knee, left ankle, left foot, right foot, right knee, and right ankle, respectively. In

Predictors | Coefficient | SE | T-value | |
---|---|---|---|---|

Constant | −18.88 | 2.981 | −6.33 | 0.000 |

_{r} |
47.610 | 21.51 | 2.210 | 0.031 |

_{r} |
−0.520 | 23.17 | −0.020 | 0.024 |

_{r} |
7.050 | 10.84 | 0.650 | 0.018 |

_{l} |
−24.34 | 46.02 | −0.530 | 0.043 |

_{l} |
71.81 | 54.37 | 1.320 | 0.032 |

_{l} |
−65.37 | 50.00 | −1.310 | 0.047 |

_{r} |
44.93 | 30.75 | 1.460 | 0.032 |

_{r} |
−4.190 | 7.440 | −0.070 | 0.044 |

_{r} |
−54.30 | 53.50 | −1.010 | 0.041 |

To test the significance of more than two joints simultaneously, ANOVA technique is used. From

S.O.V | DF | SS | MS | F-value | |
---|---|---|---|---|---|

Regression | 9 | 1368.87 | 152.10 | 26.14 | 0.000 |

Errors | 56 | 325.79 | 5.82 | - | - |

Total | 65 | 1694.67 | - | - | - |

Where, DF is the degree of freedom and SS is the sum of square for regression, MS is the mean square, and F-value is the calculated value of F-statistic. In ^{2} and Adj-R^{2}, which show higher accuracy of 88% and 91%, respectively. It means that only nine joints play significant role in the age estimation of a walking person.

Model | R^{2} |
Adj-R^{2} |
---|---|---|

Sh_{r}, E_{r}, W_{r}, H_{r}, Hip_{r}, K_{r}, A_{r}, F_{r} |
70% | 62% |

Sh_{l}, E_{l}, W_{l}, H_{l}, Hip_{l}, K_{l}, A_{l}, F_{l} |
60% | 54% |

Hip_{c}, Spine, Sh_{r}, Head |
58% | 48% |

Sh_{r}, E_{r}, H_{r}, K_{l}, A_{l} , F_{l}, K_{r}, A_{r}, F_{r} |
91% | 98% |

The screen shorts of experiments 1, 2, 3 and 4 are shown in

The difference between observed age y_{i} and estimated age

y_{i} − |
y_{i} − |
y_{i} − |
y_{i} − |
Other value |
---|---|---|---|---|

49 | 00 | 01 | 00 | 00 |

There are forty-nine observations for which the difference is zero between the observed value and the estimated age of a walking person. On the other hand, there is one observation for which the difference is 2. The results indicate the applicability of the model for age estimation of the walking persons.

The relationship between the estimated ages of a walking person with the most-contributed joints is shown in

The model shows a linear relationship between estimated ages,

From

Similarly, there is a linear and direct relationship between estimated age and the right hand joint during walking. As the average value of the right hand joint, increases with the increase in age of a walking person. Also, the observations are clustered around the fitted line and show less variation as compared to other joints. There is a linear and direct relationship between the estimated age and all the joints under consideration. This means that the average value of joints increases with the increase in age of a walking person.

This research work aims to calculate the age of a walking person on the basis of the 3D Kinect and general linear regression models. First, the authors have considered all the possible joints for the age prediction of walkers. With the help of the best model selection criteria R^{2} and adjusted R^{2}, the authors then used the most significant or contributing joints of a walking person for the estimation of age. The highly contributed joints are shoulder right, hand right, elbow right, knee right, knee left, ankle right, ankle left, foot left and right, which predict the age of a walking person. ^{2}) for the proposed model is 91% and the adjusted R^{2} is 88% (see ^{2} is high, which is decided on the basis of R^{2}. As the

We compared the proposed system with the existing approaches of gait-based age classification according to dataset, technique, and accuracy. All the existing approaches generated their own dataset, but Zhang et al. [

Author | Years | Data set | Technique | Accuracy |
---|---|---|---|---|

Davis [ |
2001 | Self (15) | Linear perception | 95.0% |

Aderinola et al. [ |
2010 | Self (14) | HMM | 83.3% |

Makihara et al. [ |
2011 | Self (168) | KNN | 80.0% |

Abirami et al. [ |
2013 | Self (178) | KNN | 86.0% |

Zhang et al. [ |
2015 | Self (26) | Random forest | 95.0% |

Yoo et al. [ |
2017 | Self (205) | SVM | 85.6% |

Zhu et al. [ |
2021 | Self (154) | Random forest | 96.0% |

The proposed method presented a large dataset with 273 subjects. The maximum accuracy in the existing approaches is 96%, while the proposed system shows 98.0% accuracy. Therefore, the proposed system is efficient in terms of accuracy and the dataset used.

This research paper proposes a novel method for age detection via gait-based 3D joint position. Our main contribution is that the age of walking subject is automatically detected in real-time via MS Kinect. The second contribution is the collection of a unique dataset of 3D skeleton joint position data, which is available on request. On the basis of a dataset generated from 273 people ranging in age from 7 to 70 years, the best age detection performance of 98.0% was achieved using CLRM, and 9 selected features (right shoulder, right elbow, right hand, left knee, right knee, right ankle, left ankle, left foot and right foot). The experimental result of the proposed approach proved that the proposed system was different from the existing approaches in terms of large datasets used, limited features, and high accuracy. Moreover, in future, the security measures can be further improved through intensive investigations of both gender and age of the walker with a single and limited set of joints. The study recommends the use of the same model for the exact identification of the person in future works apart from the recognition of age and gender only.

The authors acknowledge the support from the Deanship of Scientific Research, King Khalid University for funding this work through Large Groups RGP.2/212/1443.

The authors extend their appreciation to the Deanship of Scientific Research at

The authors declare that they have no conflicts of interest to report regarding the present study.