Elsevier

Gait & Posture

Volume 47, June 2016, Pages 80-85
Gait & Posture

Accuracy of KinectOne to quantify kinematics of the upper body

https://doi.org/10.1016/j.gaitpost.2016.04.004Get rights and content

Highlights

  • A marker-less motion capture system was compared to a standard marker-based system.

  • Functional upper extremity movements were assessed.

  • Marker-less system with sufficient accuracy for clinical setting.

  • Functional movements should be performed in sitting.

Abstract

Motion analysis systems deliver quantitative information, e.g. on the progress of rehabilitation programs aimed at improving range of motion. Markerless systems are of interest for clinical application because they are low-cost and easy to use. The first generation of the Kinect™ sensor showed promising results in validity assessment compared to an established marker-based system. However, no literature is available on the validity of the new ‘Kinect™ for Xbox one’ (KinectOne) in tracking upper body motion. Consequently, this study was conducted to analyze the accuracy and reliability of the KinectOne in tracking upper body motion.

Twenty subjects performed shoulder abduction in frontal and scapula plane, flexion, external rotation and horizontal flexion in two conditions (sitting and standing). Arm and trunk motion were analyzed using the KinectOne and compared to a marker-based system. Comparisons were made using Bland Altman statistics and Coefficient of Multiple Correlation.

On average, differences between systems of 3.9 ± 4.0° and 0.1 ± 3.8° were found for arm and trunk motion, respectively. Correlation was higher for the arm than for the trunk motion.

Based on the observed bias, the accuracy of the KinectOne was found to be adequate to measure arm motion in a clinical setting. Although trunk motion showed a very low absolute bias between the two systems, the KinectOne was not able to track small changes over time. Before the KinectOne can find clinical application, further research is required analyzing whether validity can be improved using a customized tracking algorithm or other sensor placement, and to analyze test–retest reliability.

Introduction

The quantitative description of human motion finds application in research and in clinical settings. A common approach is with marker-based systems (MBS), where markers are placed on the skin. Such systems are used widely in research laboratories and are highly accurate [1]. However, their use has disadvantages: data collection and processing are time-consuming, require highly trained personnel, and are restricted to the laboratory setting. Markerless systems have evolved alongside the technical advancement of cameras and sensors. The Kinect™ from Microsoft, which was developed to control video games through body movements, has become of interest to the research community. The Kinect™ is able to track three-dimensional motion by combining information from a color camera and a depth-sensing infrared camera. It is of particular interest for clinical settings, since it is relatively low-cost, does not require time-consuming setup, can be used in various spaces and is easy to use.

In order for the Kinect™ to be used in clinical settings from a biomechanical perspective, the system needs to have sufficient validity to measure kinematic changes. This would allow, for example, determining the reduced shoulder range of motion (ROM) of a frozen shoulder patient and the monitoring of their progress during physiotherapy on a monthly basis. To achieve this, the system needs a measurement error of ROM of less than 7.7° (flexion), 6° (abduction) and 3.7° (rotation) [2].

Different studies have examined the accuracy of the Kinect for tracking the human body. For shoulder abduction in the frontal plane, a good correlation of ROM between the Kinect and a MBS was found; while for elbow flexion in the sagittal plane, a decreased correlation was obtained [3]. Accordingly, a larger bias for shoulder flexion than abduction was reported [4]. This indicates a dependability of the validity of the Kinect on the plane of motion. Generally, larger differences in kinematic measures were found for lower extremities compared to upper extremities [3], [5], [6], [7]. Clark et al. found a bias proportional to the measured value of Kinect compared to a MBS for the pelvis and sternum, but not for the hand [8], while others noticed a poorer correlation for the trunk than the shoulder angle for the Kinect compared to a MBS [4]. This shows a difference in validity between the core of the body and the extremity for the Kinect. Most studies examined accuracy in the standing position [3], [4], [8]

Most previous studies were executed with the first generation Kinect (KinectV1) [3], [4], [5], [6], [7], [8]. In 2014, the new Kinect™ for Xbox one (KinectOne) was released by Microsoft. This system is based on higher quality sensor technology (1920 × 1080 instead of 640 × 480 resolution for the color and 512 × 424 instead of 320 × 240 resolution for the depth-sensing camera), as well as an enlarged field of view compared to KinectV1. Additionally, according to the manufacturer's specification, the algorithm for motion detection has been improved.

It can be speculated that the technological improvements result in higher accuracy in body tracking and, consequently, a higher validity of KinectOne to track movements. A study has found that, generally, KinectOne has excellent concurrent validity for spatiotemporal measurements and anterior–posterior measures during dynamic and static balance tests, but consistently poor to modest validity for kinematic parameters of the lower body and medial-lateral measures during balance tests [9], [10]. Therefore, the aim of this study was to determine the concurrent validity and intra-session reliability of the KinectOne compared to a MBS for measuring segment angles of the trunk and upper extremities during functional movements.

Section snippets

Methods

Twenty subjects participated (age: mean ± SD: 33 ± 9 years; height: 173.7 ± 8.4 cm; weight 65.9 ± 10.6 kg; 10 female) and signed informed written consent. The study was approved by the local ethics committee. The subjects wore tight-fitting shorts (women with bra). Before data collection, each subject was equipped with 39 reflective markers, according to the plug-in-gait full body model [11]. Data were simultaneously collected using a 6-camera Vicon System (200 Hz, VICON, UK) and the KinectOne (30 Hz,

Results

On average, between-system differences of 3.9 ± 4.0° and 0.1 ± 3.8° were found for arm and trunk motion, respectively. For inclination and rotation exercises the direction of bias for the arm was positive (overestimation). Contrary, KinectOne overestimated inclination of the trunk (2.4 ± 2.8°) but underestimated rotation (−3.3 ± 2.0°). RC was found to be smaller than the range from lower to upper LoA for both segments in all exercises (Table 1). Results of system comparisons are shown in Table 1

Discussion

To determine the accuracy of KinectOne in tracking human motion we analyzed concurrent validity and intra-session reliability of KinectOne and a MBS. The motion of the arm and the trunk were recorded in different planes of motion while sitting and standing using both systems simultaneously.

Literature reports lower accuracy of KinectV1 in tracking trunk motion compared to motion of the upper extremities [4], [8]. Our data showed the opposite result for the absolute bias. However, we have to

Conclusion

The results of this study revealed that the accuracy of KinectOne in tracking arm motion is sufficient for clinical settings, with the exception of standing shoulder flexion. The recommendation is that the movements be performed seated. Although absolute bias of trunk motion was generally smaller, KinectOne is not able to track small changes in trunk motion due to the high RC/SRD and low CMC. Future research is needed to improve tracking of the trunk, and to establish whether a different

Conflict of interest statement

None of the authors have any financial or personal relationship with other people or organizations that could inappropriately influence their work.

Acknowledgments

The authors would like to acknowledge the assistance of Mariella Oswald for data collection, Andre Meichtry for statistical consultation, and Christian Schärli for providing software.

There was no study sponsor with influence on this study.

Cited by (0)

View full text