1. Introduction
The main aim of this study is to analyze and visually represent subtle variations of similar human motion in a structured manner. Human joint–bone motions can have numerous subtle orientation variations. Although these subtle variations are visually similar, the underlying orientations can differ markedly. Visualizing human motion with 244 kinematic degrees of freedom (DoF) is a challenging study area. Generally, existing research represents user poses and positions over time in order to visualize human body motion in 3D [
1]. Together with spatial and temporal data, object orientation in 3D produces extensive data sets that require processing, organizing, and storing for 3D reconstruction.
This paper proposes Motion-sphere for visualizing human motion using data from multiple inertial measurement units (IMUs). Most contemporary motion capture studies record human joint orientations using sensors and/or camera and then apply them directly to a 3D avatar for reconstruction [
2,
3]. However, there is a considerable requirement for the accurate and correct visualization of human motion at the minute scale, offering insights into and visual differentiation of subtle changes in intended human motions.
It has become common practice to use IMUs for motion capture and analysis, providing more precise orientation estimates for any rigid body (in this case, human bones) as compared to camera-based or any other form of motion capture system [
4]. Consequently, commercial vendors commonly offer hardware support for filtering noisy data, software compensation for heading drift, multiple IMU time synchronization, etc. [
5]. Rigid body orientations that are calculated by sensors are presented as quaternions (four-dimensional (4D) numbers) that are widely used to realize three-dimensional (3D) rotation in animation and inertial navigation systems. Quaternions can communicate multiple orientation changes as a single change without causing gimbal lock situations. Hence, we adopted quaternions as the basic mathematical unit to represent human body orientation in 3D, consistent with most human motion analysis research [
6,
7,
8].
Any human motion is a sequence of quaternion rotations, visually represented as a trajectory on a 3D unit sphere.
Figure 1 shows an overview of the proposed Motion-sphere. A 3D LiDAR sensor is used to track the position of a moving human body. The 3D position of the pelvis is computed while using the segmentation of raw point cloud data based on the user’s height and IMU sensors are used to estimate the orientation of each bone joints during human body motion in a real environment. When a user performs a motion, orientation data from the IMU sensors are transformed into a 3D avatar’s coordinate frame for reconstruction.
The quaternion frames are decomposed into swing and twist [
9] rotations and visualized as trajectories on a unit sphere. Mapping a grid on the unit sphere helps us to quantify specific human motion and individual bone movements and their respective rotations can be visually analyzed. The Motion-sphere enables comparisons between two sets with visually similar motion patterns. Therefore, the technique opens up itself to a wide range of applications, including athletic performance analysis, surveillance, man-machine interfaces, human motion tracking, kinesiology, physical fitness training, and therapy.
Each subtle movement is mapped on the unit sphere as a series of 3D points rotating about given X, Y, and Z axes. For example, lower limb movement in a sit-up action would be represented as a trajectory on a unit sphere. Although the left and right legs have similar movements, the subtle variation in orientation could not be identified in a video. Likewise, a sit-up performed a second time may have a different trajectory to the first. The human motion details are contained within the axis and orientation changes reflected by the quaternion trajectory. This paper visualizes human motion while using this trajectory, thus providing an excellent candidate for use in the aforementioned applications.
The current work validates the proposed Motion-sphere by considering the range of motion (ROM) from the TotalCapture dataset [
10]. The reconstruction accuracy, visualization accuracy, learnability, and usability of the Motion-sphere are validated against a few existing motion visualization methods. We have demonstrated the use of Motion-sphere for comparing multiple human motion, authoring, and editing, and as a standard to validate the accuracy of reconstruction.
The remainder of this paper is organized, as follows.
Section 2 discusses related work with respect to various methods used for human motion visualization.
Section 3 and
Section 4 present the hierarchical joint–bone segment and discusses details on decomposition and visualization.
Section 5 validates the Motion-sphere and
Section 6 demonstrates a few key applications of the Motion-sphere.
2. Related Work
Visualizing time-varying positions and orientations of multiple human joints is challenging. Among the various techniques proposed thus far, visualizing motion in two-dimensional (2D) image is widely practiced and considered to be the most intuitive approach [
11]. Variations of hand gestures are visualized on a 2D strip using a stick model [
12]. Although other metrics are considered to compare the gesture, orientation changes in the axis parallel to the torso (vertical axis) cannot be noted. Thereby, the temporal dimension of the human motion is fully overlooked.
Motion-cues is yet another visualization technique using key-poses [
13]. They focus on the motion of a specific bone segment using directed arrows, noise waves, and stroboscopic illustrations. The primary focus is the artistic depiction of motion rather than the subtlety of time, space, and orientation. Such visualization is more appropriate for creating artistic or multimedia work [
14]. Motion-track [
8] uses a Self-organizing map to construct a sequence of key-poses that it connects through a 2D curve. This kind of visualization is suitable for full-body motions, like walking, running, jumping, or marching. Although it uses hierarchical skeletal model, there is no evidence of any visual depiction of individual bone segments orientations. Zhang et al. [
15] proposes a technique of generating motion-sculpture (MoSculp) from a video. Again, this is useful for creating artistic expressions of human motion; it expresses motion as a 3D geometric shape in space and time. The orientation of joints are overlooked in the MoSculp, which makes it difficult to reconstruct the motion or sculpture.
Among various visualization methods, Oshita. et al. proposed a self-training system (STS) for visually comparing human motion by overlapping 3D avatar models [
16]. The authors considered the motion data of tennis shots played by an expert vs. a trainee to compare the motion in terms of spatial, orientation, and temporal differences while using regular and curved arrows for visualization. Because 3D avatars are used for visualization, the technique does not exaggerate the differences and it becomes unnatural and difficult to note subtle variations in orientations.
Yasuda et al. [
17] proposed a technique, called the Motion-belt, which visualizes motion as a sequence of key-poses. It offers temporal information on motion by varying the distance of the key-poses on a 2D belt. Specific color coding is used to represent the orientation of the pelvis (Full Body) in 360°. Although Motion-belt can successfully provide an overview of the motion type at a higher level of abstraction, it neglects details, such as the orientations of individual bones, and it is challenging to distinguish between human motions with subtle orientation variations. In addition, comparing a multitude of variations with respect to their orientation and position over time is highly difficult.
It is easy and intuitive to represent orientation while using a quaternion. Hanson [
18] described a concise visualization of quaternions, including representing points on the edge of a circle, hypersphere (
), single quaternion frame, etc., with several useful demonstrations, such as the belt trick, rolling ball, and locking gimbals. The approach is mainly for educational and demonstration purposes. Quaternion maps are the most common techniques for visualizing quaternions with the help of four spheres (
). Each of these four spheres has a particular meaning, and visualizes the raw quaternions as a combination of three tuples ([x,y,z], [w,x,y], [w,y,z], [w,x,z]) as 3D points on a unit sphere, collectively as a trajectory. However, for human motion visualization, quaternion maps do not provide distinctive visual differences between human motion trajectories with subtle differences.
In one of our previous prototype studies [
4], we visualized human motion in two-dimensional (2D) as a trajectory. Here, the raw quaternion data was considered, and a similar approach was adopted as in the current work for computing the 3D-points. A UV mapping technique [
19] was used to map the 3D-points onto a 2D grid. The hierarchical nature of human bone segments was not considered an initial vector
in quaternion form was regarded as a starting point for any human bone motion visualization. The visualization did not consider the decomposition of quaternions into twist and swing rotations. Therefore, quantifying human motion was not an option. In another work [
20], we decomposed the rotations into twist and swings and represented them on the unit sphere as 3D beads. The representation was accumulated rotation of all the parent segments into a single trajectory. Therefore, the kinematic hierarchy was not maintained in the details of the trajectory.
Various motion-databases, including CMU Mocap [
21], MocapClub [
22], Kitchen Capture [
23], CGSpeed [
24], and so on are publicly available online data sets. These represent major hierarchical rotation, scaling, and positional data of body segments in the form of Euler angles or quaternions. The list of motions they have captured includes walking, climbing, playing, dancing, sports performances, and other everyday movements. The current work shows examples of the visualization of ROM from the TotalCapture dataset [
10] for demonstration and comparison. The TotalCapture dataset provides sequences of orientation data in the form of quaternions, and the position of joints in 3D with respect to the global reference frame. We validate the reconstruction accuracy of the avatar while using the same dataset.
4. Demonstration of Trajectory Patterns
Figure 7 shows the trajectory of a simple right upper arm rotation from the TotalCapture dataset [
10]. The upper arm takes a full rotational swing in the anti-clockwise direction with a twist of almost 120°. In addition, the rotational swing is clearly on the right of the Motion-sphere, as indicated on the sphere’s grid texture. The black arrows indicate the direction of the swing.
The Motion-sphere always visualizes the bone-segment motion with respect to the parent bone-segment’s orientation.
Figure 8 shows a ROM movement of a right upper leg, in which the right upper (parent) and lower (child) legs appear to be moving in opposite directions in the initial phase (key frames 1–2). Although the orientation of the lower leg does not change significantly, it seems that the lower leg is rotating to the back of the upper leg (purely due to the upper leg rotation). In the second phase (key frames 2–3), the lower leg exhibits no movement (
Figure 8c) w.r.t the upper leg, while the upper leg changes laterally by 40° right (
Figure 8b). An approximately 80–85° twist can be noted due to the color change in the miniature upper leg model due to the lateral movement of the upper leg toward right.
Figure 9 shows another twist-swing example that visulaizes the right arm swing. Key frame 1–2 are in a 90° upper arm swing movement toward right side of the body. A twist of 90–100° is noted from the color change due to the lateral swing. The lower arm swings in front w.r.t. the upper arm by 140° up and 10° right. The color change is minimal, with approx. 40° twist.
Figure 9b shows the visualization.
Figure 10 visualizes the rotational movement exhibited by pelvis (
) and the chest (
). The corresponding spheres that are associated with
&
are selected among the menu options for visualization. Unlike the limbs, where the trajectories start at the south pole of the sphere in the attention pose, the trajectories for the pelvis and chest begin at the north pole, as shown in
Figure 10. A 70° bend in
and 60° bend in
(w.r.t.
) is seen.
7. Discussion
Generally, people find it comfortable to follow a video for fitness training, as it is a common medium used to observe any kind of human motion. However, Motion-sphere is more informative for detecting patterns in user motion, so that users can imitate experts’ movements more accurately. In the case of video clips, there is no definitive way to quantify a twist or pattern of swing as we discussed in view of several studies in the previous section. In addition, a subtle twist or swing in a bone–joint segment can be completely overlooked due to the 2D nature of the video clip. We visualized various human motion data from the TotalCapture data set and tested the Motion-sphere for a yoga training application, with a sequence of yoga movements, called the Surya Namaskara. From the trainer’s perspective, the Motion-sphere supports them to provide distinct corrective suggestions to their pupils regarding their body posture. Motion-sphere is effective for analyzing and visualizing the structure and relationship between the joint–bone segments and the effect of a parent joint on the child. Every person has a different skeletal anatomy in terms of bone length and flexibility, which are captured during calibration (
Section 3.1). However, the captured motion data in a single calibration session is more suitable for comparative visualization
The kinematic hierarchy that is discussed in
Section 3.1 covers 10 bone-segments, including the pelvis (
), chest (
), upper limbs (
,
,
, &
), and lower limbs (
,
,
, &
). The neck and shoulder bones move depending on the chest, while the hip bones move depending on the pelvis. Additionally, various other bones, like the feet, hands, fingers, and head, are merely extensions in the hierarchy by either increasing the number of sensors or by adopting deep learning methods like in the [
28,
29]. The Motion-sphere could be extended to these bone segments, depending on the application. The future work intends to extend Motion-sphere for the aforementioned bone-segments.
Each bone-segment movement is mapped and visualized on a unit sphere as trajectory, as discussed in
Section 3.2. It is natural that the trajectories overlap and clutter for a prolonged motion (constitutes multiple key poses), leading to difficulty in understanding the motion. Therefore, the Motion-sphere works best when the movement between lesser number of key poses are visualized. This also compliments the very purpose of visualizing the subtle motions while using the Motion-sphere.
8. Conclusions
This paper focused on the visualization of subtle human motion. Visualization helps to quantify individual joint bone movements for comparative analysis. We used quaternions for 3D rotations. Visualizing and recognizing subtle human motion variations greatly enhances motion capture and reconstruction quality, both of which are essential for regular usage and research. Among other benefits, the current work enables opportunities in fitness, rehabilitation, posture correction, and training. Future work will develop the current work into a standard for the visual representation and analysis of human motion.
Most work in the area of human motion visualization is based on 3D avatar analysis. Motion-sphere is a novel approach for recognizing and analyzing subtle human movement based on motion trajectory visualization. As every single orientation of human motion has a unique representation on a Motion-sphere, it offers a wide scope in the future to alter or edit motions to eventually generate new variations. The positional data of the user from the LiDAR helps in realistic reconstruction of the performed motion. In the future, we intend to extend the scope of our research to visualizing positional data. The position of human joints is strictly hierarchical and depends on the position of the pelvis. Therefore, exploring ways of visualizing the pelvis’s position in 3D space would be interesting to see alongside the Motion-sphere. Thus, the scope of Motion-sphere would not be limited to joint–bone analysis, but it can extend to areas, such as locomotion, and gait analysis. The authoring of human motion or editing captured motion from the sensor seems like a very interesting application of the Motion-sphere. Therefore, exploring user-friendly ways of authoring human motion using the Motion-sphere and standardizing the Motion-sphere for avatar reconstruction validation are promising future works.