A Review on the Use of Microsoft Kinect for Gait Abnormality and Postural Disorder Assessment

Gait and posture studies have gained much prominence among researchers and have attracted the interest of clinicians. The ability to detect gait abnormality and posture disorder plays a crucial role in the diagnosis and treatment of some diseases. Microsoft Kinect is presented as a noninvasive sensor essential for medical diagnostic and therapeutic purposes. There are currently no relevant studies that attempt to summarise the existing literature on gait and posture abnormalities using Kinect technology. The purpose of this study is to critically evaluate the existing research on gait and posture abnormalities using the Kinect sensor as the main diagnostic tool. Our studies search identified 458 for gait abnormality, 283 for posture disorder of which 26 studies were included for gait abnormality, and 13 for posture. The results indicate that Kinect sensor is a useful tool for the assessment of kinematic features. In conclusion, Microsoft Kinect sensor is presented as a useful tool for gait abnormality, postural disorder analysis, and physiotherapy. It can also help track the progress of patients who are undergoing rehabilitation.


Introduction
Microsoft's Kinect sensor is a motion-sensing device that gives users the features to interact with game consoles and computers via ways such as gestures, spoken commands, or movement [1]. Kinect sensors provide new and enhanced features for motion detection and 3D reconstruction. Kinect sensors also introduce many features that allow for more accurate research into the movement of the human body and its gestures. e sensors allow for interaction through voice commands that is a unique component of the technology. It has detectors and infrared emitters to capture human physical activities.
e key components of Microsoft Kinect sensors are the RGB cameras, IR depth sensors, and the multiple microphone array. e second version of Kinect has some enhanced features compared to earlier Kinect [2]. e colour camera of Kinect v2 has 1,920 × 1,080 @30fps while that of Kinect v1 has 640 × 480 @30fps. In terms of the depth camera capabilities, Kinect v2 uses 512 × 424 pixels, while Kinect v1 uses 320 × 240 pixels, and as a result, Kinect v2 has better image recognition compared to the earlier version. Kinect v2 is noted to have a wider area view compared to Kinect v1. Another key feature is that Kinect v2 has better skeletal joint tracking where it is able to capture 26 joints, whereas Kinect v1 can only capture 20 joints. e unique feature of Kinect sensors can be applied to the medical field for the purposes of diagnosing diseases and physiotherapy rehabilitation of people who may have walking disabilities due to physical injury or related diseases.
As stated above, Kinect has found application in many areas related to posture and motion capturing. e major bulk of studies are related to Kinect research in the areas of motion tracking, monitoring, diagnosis, and rehabilitation. Some representative studies with Kinect technology include: Lavanya et al. [3] presented dynamic finger gestures with skeletal data extracted from the depth sensor. A unique technique was designed for the recognition of dynamic gestures that can be used in auditoriums and classrooms.
is approach allowed for more dynamic hand gestures to be developed that can be used in different environments. An example is a tutor using this technique to instruct students in a classroom who have speech problems to assist in their studies. e use of Kinect for medical monitoring and diagnosis has also been trialed by researchers. Ales Prochazka et al. [4] presented a novel technique of using Kinect for heart rate estimation and breath monitoring to determine the likelihood of any medical condition. e mean thorax movement was monitored within a selected area to estimate the breathing of patients. Huy-Hieu et al. [5] presented a realtime system for the detection of objects for patients who are visually impaired. A unique system was designed that allowed visually impaired people to move freely and to detect any obstructions. Object detection was based on the 3D information captured with a depth sensor. However, the designed system was limited to only indoor use. Xin Dang et al. [6] presented a novel interactive system with an electroencephalogram and depth sensor for people with dementia. Skeletal data captured from the depth sensor were extracted to determine the motion of a user and their mental state. e designed system using a deep neural network can be used to aid patients with dementia. Torres et al. [7] provided a novel approach to assist physicians in the diagnosis of Parkinson's disease using posture and movement captured with Kinect. e characteristics of movements such as frequency and amplitude were essential to study tremors in people with Parkinson's. e results achieved in the study can assist clinicians to diagnose Parkinson's based on the tremors intensity and the postural changes.
Kinect sensors are also widely used for the purposes of rehabilitation. Capecci et al. [8] demonstrated an innovative approach in the evaluation of dynamic movement in a rehabilitation scenario. ey were able to track skeletal joints in evaluating the performance of patients during a low back pain physiotherapy exercise. Postolache et al. [9] developed a unique framework for physiotherapy assessment based on a mobile application using skeletal data. e designed system assists physiotherapists to improve the effectiveness of the training sessions for patients undergoing rehabilitation. Monique Wochatz et al. [10] illustrated a reliable and valid assessment of the lower extremity rehabilitation of exercises using Kinect v2 sensors. e authors demonstrated the Kinect sensor as a reliable tool in assessing the lower limb position and the joint angles during exercises. Sanjay et al. [11] developed a unique framework for stroke rehabilitation of patients in a home environment. A framework was designed to aid patients who have suffered a stroke in their treatment process. e designed system can be used indoors to help patients who have difficulty in movement.
Abnormal gait is the asymmetric movement of a person that is most likely to be caused by disease or physical injuries.
is could be a result of nerves damage, injuries, weakness of muscles, or joint problems. e detection of gait abnormalities at an early stage can help prevent other complications.
ere are traditional wearable sensors for gait abnormal detection; however, these conventional methods are quite cumbersome to use. Kinect is seen as a better alternative to wearable sensors in gait abnormal detection.
As such, gait analysis has also received wide interest among researchers [12][13][14]. Kinematic features such as the step length, walking speed, and cadence are useful to determine the gait of an individual.
is will normally be cyclical and symmetric unless there are some forms of abnormalities.
e human gait could be affected by the musculosketal and neurological systems as well as the motions habits [15]. Hassain Bari et al. [16] presented a novel method by designing a deep neural network for gait recognition. is was then evaluated with a 3D skeletal gait data set. Another study by Wan Zharfan et al. [17] illustrated an economic technique of gait analysis based on pixel coordinates of body joints. e technique served as an alternative method to determine gait parameters in a Vicon motion analysis. e human posture is the physical positioning by which the body takes at a particular time. Posture is the arrangement of the structure of the human body and its position. Correct body posture can help reduce pressure on the human body by keeping it balanced. e human body posture may be intentional or unintentional due to natural causes. ere are several techniques for postures recognition with skeletal data. Samiul Monir et al. [18] presented a novel technique with a rotation and scale-invariant for posture recognition from skeletal features. is technique for posture recognition used skeletal data and angle rotation of an individual. A set of vectors and manipulation of angles were used to determine the posture. Zequn et al. [19] developed a novel posture recognition model that can be used to identify different postures captured. e depth sensor was used to generate features of different body parts of an individual. e captured features were then fed into the support vector machine (SVM) to identify the posture.
Although there are quite a number of studies available using Kinect for the assessment of gait and posture abnormalities, to the best of our knowledge, there is no overall review study that attempts to summarise articles on the use of the Kinect sensor for gait abnormality and posture disorder assessment. Our review adopts a systematic approach similarly used by Shmuel Springer et al. [20].
e study provides up-to-date review of the articles for analysis and discussion.

Methods
In this section, we retrieve articles that meet the inclusion criteria for our study. e existing articles identified were summarised into tables indicating the methods used and the sampling area. e sample area stated the source from which the article was retrieved, the authors of the article, and the year it was published. Our focus was on articles that used Kinect sensors for assessing gait and posture abnormalities.

Search and Identification of Articles.
e scholarly database used to identify the articles were: IEEE Xplore, Sci-enceDirect, CINAHL plus, and PubMed. e search was done in two different parts as follows: (i) Part 1: the use of Kinect sensors for the detection of gait abnormality or disorder 2 Journal of Healthcare Engineering (ii) Part 2: the use of Kinect sensors for detection posture disorder or instability e key terms used in part 1 of the search were: "Kinect sensor," "gait abnormality," "gait disorder," and "walking abnormality." In the second part, the key terms used were: "Kinect sensor," "posture disorder," "posture abnormality," and "posture instability." e search was conducted between January and May 2021 to retrieve the most recent articles.

Eligibility Criteria of the Identified Articles.
In identifying the articles for the study, we conducted two different searches for the database by two independent authors. e authors were able to identify and remove duplicate articles from the various database. e elimination of irrelevant articles was done without bias or oversight in order to get all relevant articles that meet the inclusion criteria. A number of diseases that could affect gait and posture abnormalities were included. ey were: Parkinson's, ataxia, multiple scoliosis, stroke, and depression. e exclusion criteria were for articles that only discussed gait detection and postures without considering abnormality. In the second part, articles that only discuss posture assessments without examining posture disorder or deformity were also not included. Articles that used wearable sensors for gait abnormality and posture disorder detection were not included. Figure 1 is a flowchart diagram of the search methodology.

Results
e initial search and retrieval of all articles from the various database were 458 for gait and 283 for posture, all using Kinect technology. e articles were further screened to ensure they meet the eligibility criteria for this study. In the end, a total of 26 articles were included for gait abnormality and 13 articles for posture disorders. Table 1 summarises all the studies that were included for the review for gait abnormality, the journals where the articles were retrieved and the year of publications. In Table 1, the methodology describes the sampling method applied, the statistical method, and the descriptive approach used. at sampling method describes the number of participants in the study, gender, and age distribution. e disease associated with the abnormality was stated. e sampling methods from the reviewed articles were categorised as fully stated, partially, or not stated. e statistical method describes the statistical techniques that were used in the analysis of the data. e statistical methods were either sufficiently used or partially used. It describes the models and mathematical equations that were used in the analysis of skeletal data. e description method used refers to capturing of the skeletal data, processing of data, algorithms used, and the analysis of the results. It also includes the tools used in the analysis of results and a detailed discussion of the findings. Finally, the description method was either adequate or partial description. e approach used in Table 1 was also similarly adopted in Table 2 except that it summarised the various articles for identifying posture disorders using Kinect. e sampling methods, statistical methods, and the description method are also indicated in this table.
In Table 3, the details of each article included in the study were categorised into two major phases. e first phase deals with the sampling technique used in each study while the second phase describes the key gait features captured with the major findings of each study. e limitations for each study were also included in the table.
A total number of 26 articles were reviewed. Most of the articles stated the number of participants in the study except in [26,28,37,42,43,45]. e majority of the studies used Kinect v2 as the main tool for capturing skeletal data for gait abnormality assessment, while a few articles used the older Kinect v1. In [37,44], the Asus Xtion PRO was used as a gold standard with a Kinect sensor for capturing skeletal data. Most of the reviewed articles did not state the data analysis tool. However, in [22,25,29,31,35,39,40], MATLAB was explicitly stated as the main tool for data analysis. In [23], the SPSS package was used as the data analysis tool. e gait features captured were mainly the step/stride length (m), stride time (s), gait speed (m/s), gait cycle (deg), gait rhythm (m/s), and step time (s). Some of the key joint angles measured were the hip, knee, and the ankle. Various algorithms were used to train the models for gait abnormality detection.
A total number of 13 articles were included for the assessment of posture abnormality or disorder. Some of the reviewed articles [42,46,50,51] did not state the participants in the studies. In Table 4, Kinect v2 was mostly used for skeletal data capturing, except in a few studies that used Kinect v1. Nine of the reviewed articles used Kinect v2 while four used Kinect v1. Most of the studies did not state any medical condition that resulted in posture abnormality. However, in [47], Parkinson's disease was stated, and in [56], a case of chronic traumatic brain injury was present. In [50,58], patients with suspected multiple scoliosis were also assessed for posture abnormality. e majority of the reviewed articles did not state the data analysis tool except in [47,48,50] that stated MATLAB. In [58], the IBM Watson Analytics was used to analyse the data for posture abnormality for patients with suspected cases of multiple scoliosis.

Gait Abnormality or Disorder.
e purpose of this study is to review the available studies using Microsoft Kinect for the assessment of gait and posture abnormalities. e key features measured included the angles formed by leg swing, speed, and distance of each gait step. ese parameters were useful in detecting gait asymmetry in order to distinguish normal from abnormal gait. Some other components from the summarised studies were the algorithms used, the major findings of each study, and limitations.
From the reviewed articles, various methods were employed in assessing and detecting gait abnormality. e gait features captured were mainly the step/stride length (m), gait speed (m/s), gait cycle (deg), and step time (s). Some other gait features captured from the studies were angles of knee joints, ankle joints, and hip angles joints. e measured joint angles were used to train the models in detecting gait abnormalities. e measurement of joint angles helped improve the efficiency and robustness of the trained models to detect gait abnormality. ey were various algorithms used to train the models for gaits abnormality detection. Some of the algorithms used in Table 3 were machine learning algorithms. e machine learning algorithms were either supervised or unsupervised, depending on the approach used. Supervised machine learning algorithms use classifications and regressions, while unsupervised use clustering and associations to determine outliers in the data. Algorithms that were used in the designed models for gait abnormality were: Bayesian algorithm, K-nearest neighbors algorithm (KNN), convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory (RNN-LSTM), isolated forest (IF) algorithm, and one-class support vector machine (OC-SVM) algorithm. Depending on the algorithm that was used and the mythological approach, different accuracy were achieved. e Bayesian algorithm was commonly used in [15,19,23,33] for the assessment of gait abnormalities. RNN-LSTN algorithm was used in three of the studies [20,36,37,40]. Amr Elkholy et al. used unsupervised one-class support vector machine (OC-SVM) and isolated forest algorithms for abnormal detection in [31,33]. Some other algorithms such as the EDSS and MSWS were also applied in [27]. e algorithms used in the various studies cannot be compared to determine which is more efficient and robust.
is is because different methods and data sets were used to achieve the desired accuracy. e general limitation of the summarised studies has to do with the relatively small data set used. Most of the studies did not use a large data set to test the robustness of the trained model except in [21,22,25,29,33,34,44] that used large data sets. Another limitation was some gait parameters were not used in training the models for abnormal gait assessment. Some studies did not also include key joint angles in the trained model. erefore, some of the models  did not give high precision and robustness in assessing gait abnormalities. Also, the use of Kinect v1 has limited capabilities compared to Kinect v2 that has more enhanced 3D skeletal tracking capabilities. Table 4, some body features were measured to determine the posture abnormality of a person. ey were the body center mass position and the shoulder position angulation. e head position and the neck angles were also essential to detect postural instability. In [47,49], the center of body mass was used to determine the postural instability. e rapid upper limb assessment (RULA) method was commonly used to assess postural instability in [52,57]. e distance between the neck, shoulder blade, and angles was very essential in determining the abnormal posture of a person. In [52], the height, hips, and shoulder position were measured as well as the shoulder angles. In [53], the knee joints, ankle joints, the lateral joints, and interior joints were computed. e spine angle was also used in [47] to determine the abnormal posture of an individual. Different algorithms were used to determine postural disorder from the summarised studies. e algorithms that were used included pattern recognition neural algorithm, CoVNet model, Berg balance scale (BBS) method, and the RULA technique. ese algorithms were used to track the static body position features and key joint angles to determine the postural instability of a person. e accuracy in the detection of posture disorder was not considered because the studies focused only on determining if there was posture abnormality.

Posture Abnormality or Disorder. From
e limitations of the various summarised studies largely depended on the small data set used in the various studies.

Mathematical Analysis of Gait Abnormality.
In Sun Bie et al. [21], the spatial position was used with the associated joints for each subject walking. e extracted joint angle formed were then calculated. e equations used in the joint angle computation were given by the following equation: where J[i] represents the joints, len (i,k) represents the distance between joint i, and k θ (i, k, j) describes the angle formed by joint i, k, and j. erefore, the angles formed by the left leg and the right leg were calculated by the following equation: where θ l and θ r represents the angles formed on the left and right leg, respectively, of a subject walking and E( * ) represents the expected value of the operation for the simulation. e equations were used in computing the key joint angles to determine gait asymmetry. e angles calculated were the hip angle, knee angle, and the ankle angle. e challenge with this technique is that there may be some difficulties in measuring the inner joint angles of subjects. e gait cycles were computed and given by the following equation: where | * | is the absolute value from the operation and 0.0333 is the conversion factor from the Kinect sensor.
In [37], the gait energy image (GEI) was used based on the Gaussian mixture model (GMM) for each pixel in the simulation. e GEI was the image captured with the Kinect sensor of an individual in a walkway. It can be used to determine the dynamic information of a gait sequence. e gait cycle was then extracted and computed at the point where a normalised autocorrelation from the silhouette image was high in the GEI.
where C(N) represents the autocorrelation for N frameshift, and the N value is chosen to empirically represent all the abnormal gait cycles that exist in the tested data sets; K � N Total -N-1 where N Total represents the total number of frames sequence. S(x, y, n) indicates the pixel values at a position (x,y) in the silhouette frame n. e GEI was then computed as an average of the normalised and aligned silhouette over the gait cycle in the following equation: where n cycle represents the frame of the gait cycle while S i is the silhouette frame of x, y pixel coordinates of the image captured. e extracted gait energy image (GEI) and the gait cycle were used to determine gait abnormality from different viewing points based on the colour image sequence. However, this technique does not consider factors that may affect the colour image sequence such as clothes variations.
In [32], the Euclidian distance was used to compute the joints and dynamic body parts for a subject to determine gait abnormality. e Euclidian distance is defined in the following equation, where the distance r and s are the shortest distance between the line segment rs: Hero's formula was then used to calculate the triangular area of the gait cycle. e area obtained by Hero's formula is given by the following equation: e step length was then calculated between the right foot from left foot in the Euclidian distance as follows: where A and B are defined as the area formed by the joint angles. e areas and angles were formed between the hip right foot and the left foot. A triangle was generated to get the area and angle between the right foot. erefore, the Euclidean distance to compute for the maximum foot distance lifted from the ground is given by the following equation: Ground clearance � max right foot y − min right foot y . (9) e limitations in using the Euclidean distance for computation has to do with the multiple dimensions and the sparse nature of data. is presents some variations in trying to measure the gait distances for subjects in a walkway.
In [44], asymmetry features were used to detect walking abnormality in a subject. e motion asymmetry between the right body parts and the left body parts of the skeletal data was extracted. e average distance extracted from the skeletal data for a pair of joint angles was then computed. e Euclidean distance used to represent the asymmetry feature was calculated from the following equation: where D ρ represents the left and right of the average distance of a subject while x it , y it , and z it represents the 3D coordinates of the joint i of a subject of frame t and n is the number of frames of action sequence. N p is the set of joints for the left and right body parts of an individual. e velocity magnitude feature was computed in the study to detect slow action performed by the subject. e equation used to calculate is as follows: Equation (11) is essential in computing the displacement magnitude for each body joint between two successive frames where N represents the number of joints and n represents the number of frames.
In [27], gait assessment of a patient was evaluated by extracting the time series for the knee angles and the gait cycle of dynamic time warping (DTW). e knee DTW distance and the hip were then calculated and averaged to get the mean DTW distance of individual patients. e mean DTW distance for the hip joints and the knee that are denoted by D HP and D KP are defined by the following equations: DTW θ RH i,t , ϕ RH q,r ⎤ ⎦ . ⎡ ⎢ ⎢ ⎣ (13)

Equations for Postural Abnormality Assessment.
Several researchers have proposed different techniques for the recognition of the human posture and 3D human reconstruction. e posture probability density is used to reconstruct the posture of human beings [60]. It is based on human body measurement that can be used to determine posture. is is used as a density estimator in the following equation: where K(.) is the kernel, h is the bandwidth of the kernel, and d is the degree of freedom of the data. e probability density estimate is then given by the function where the weights c in each kernel are based on the reduced set density estimation (RSDE). e RULA-based method is a common technique for posture assessment of an individual. is technique can be used to determine the postural abnormality of an individual. Two techniques are used to calculate the joint angles, which are input from a module score. ese techniques use a voxelbased angle estimation in which the RULA score for the upper joints is computed based on their location. e joint angles are computed using vectors that are dependable on the location of each joint with correspondence to the parent joint location. is is given by the angle between two vectors of the parents' joints and the child's joints in the following equation: e magnitude of the two vectors P 1 and P 2 are calculated by: , where the computed value is then submitted into equation (15). e RULA method is a good technique for posture assessment because it is easy and fast to use. is can be used in the evaluation of posture disorder without the need to conduct any experimental measurements. is technique is, therefore, significant to conduct risks of musculoskeletal disease with regard to the posture of an individual.

Conclusion
In this study, we presented Microsoft Kinect as a noncontact tool for the assessment of gait abnormality and posture disorder. While there are several studies on gait recognition, only a few have dealt with the assessment of gait and posture abnormalities. Early detection of gaits and posture abnormalities plays a significant role for clinicians to provide corrective rehabilitation measures. Even though this is a comprehensive study, there may be some articles that are not included.
In our study, we presented 26 studies for gait abnormality assessment and 13 articles for posture disorder. e summarised studies differ by the methodology used, the gait features extracted, and the analytical tools used to process the skeletal data. Different algorithms were applied in the summarised studies, and some of them made use of machine learning algorithms. e results showed what has been done so far in the area of gait and posture abnormality assessment.
From our analysis, Kinect sensors have a high success rate of approximately 87% in abnormalities assessment. It has an accuracy ranging between 83% and 98.1% from the summarised articles for gait abnormality. is is quite acceptable in the clinical settings for the purposes of diagnosis of diseases associated with gait and posture disorders. Although Microsoft has stopped the release of Kinect sensors, it is still an important tool for diagnostic purposes. It can be concluded that Kinect sensor is an essential monitoring tool for use in medical diagnostics and can also help track the progress of patients who are undergoing rehabilitation.
Data Availability e data that support the findings of this review paper can be sourced from the summarised table in the study and the references provided.