Development of Upper-Extremity IMU to OpenSim Data Conversion Method
Team Members
Elizabeth Alderman
Kevin Brenner
Caitlin Hogan
Time Line
Project Video
Background and Motivation
Two common approaches for experimentally tracking motion are Visual Motion Capture (MOCAP) and Inertial Measurement Units (IMUs). Using MOCAP, markers on the subject are visually tracked and the resulting positions can be differentiated into joint angle velocities and accelerations. Using IMUs, point accelerations, as well as angular velocity and magnetic orientation, are collected and can be integrated into joint kinematics. While MOCAP is widely used, its applications are restricted to lab settings and are costly to implement. In contrast, an IMU's relatively small size and use of stand-alone sensors makes the technology ideal for 'in the wild" motion capture. Many industries would benefit from the use of IMU technology, including rehabilitation medicine, sports performance, robotics, and entertainment. Unfortunately, there is no widespread consensus on IMU placement or calibration, leading to much confusion in the field.
In order to accurately capture joint kinematics from the IMU data, calibration must be used to map the sensors to the anatomical segments and the placement of the IMUs must be known. The benefits and drawbacks of different calibration methods and IMU placements is a widely researched topic. Bouvier 2015 investigated the accuracy of three different IMU calibration methods to relate sensors to segments. This paper served as the basis for our investigation; however, we feel that there are still additional concerns to address. A large drawback of the Bouvier study is the reliance on the XSens algorithm to obtain the motion kinematics. This creates a black box in the experimental process, preventing the researcher from understanding what assumptions are being made and blocks the user from adjusting the equations to fit specific applications. While this research into calibration methods is valuable, the 'black box' brings many unknowns into the picture, making it impossible to fully understand why, and even if, one calibration method is better than another. It also greatly limits the types of calibration that can be performed and effectively locks the IMU placements to match those the black box expects, unnecessarily limiting the potential uses for IMUs.
The goal of this project is to develop a transparent, open source algorithm which will replace the 'black box' algorithms used in previous studies. The algorithm takes experimental upper extremity IMU data and performs the necessary calibrations and modifications to transform it into an OpenSim compatible format. From there the data will be mapped to a model to perform kinematic analysis.
Practical Goal and Related Questions
Practical Goal: Develop a program to import Upper Extremity XSens IMU data into OpenSim for kinematic analysis.
Related Questions:
How does different "ground" IMU placement affect IK accuracy?
How does "ground" IMU tilt affect IK accuracy?
Of the available upper extremity models, which provides the most accurate simulation from the IMU data and why?
Can answers to the above questions inform a more accurate IMU placement or calibration method? If so, what are the recommended IMU placements and calibration methods?
Methods
1. Data Collection Method
1.1 IMU Placement
Data was collected using three IMUs placed on the right arm and three on the torso as shown in the figure below. The IMUs on the arm were located on the hand, mid-forearm, and mid-upper arm. The arm IMUs were aligned in the plane of the arm to the best visual estimate of the underlying bone's angle, placed on the lateral side of the arm. The provided straps were wrapped securely around the subject's upper forearm and proximal upper arm. Provided hand covers were used for IMU attachment at the hand.
The IMUs on the torso served as the 'ground' IMU when processing the data. Due to the importance of the ground IMU position on the results, three different ground IMU placements were evaluated. In order to test multiple ground placements on the same motion, data was collected with all three torso IMUs active. The data was then processed by selecting only one torso IMU to be the ground IMU. The data from the other torso IMUs would not be accounted for in that simulation. This was repeated for each IMU on the torso. The evaluated ground IMU placements were the stomach, lower back, and upper back. The stomach and lower back IMU were secured with a strap wrapped around the subject's waist. The upper back IMU was placed on a strap wrapped around the subject's chest at the largest circumferential point.
1.2 Motions Captured
Three motions were captured for kinematic analysis: elbow flexion, shoulder flexion, and wrist supination. The motions were chosen from the Fugl-Meyer Assessment, a method used for motor recovery evaluation after stroke. These motions were chosen because they effectively isolated the three joints of interest. Additionally, these motions reflect those which patients would perform if IMUs were to be used for quantitative Fugl-Meyer assessment. GIFs of the motions being performed can be seen in the videos below.
2. Models Evaluated
Two OpenSim models were evaluated for accuracy with the developed algorithm. MoBL-ARMS Dynamic Upper Limb is an upper limb model created by Saul et al. (https://simtk.org/frs/?group_id=657). This model has 7 degrees of freedom and models all muscles in the right limb. The second model evaluated was developed by Seth et al. (https://simtk.org/projects/scapulothoracic) It models the Scapulothoracic Joint and has17 degrees of freedom, although no muscles are included.
2.1 Initial Pose of the Model
The initial pose of the model must match the pose of the subject to ensure accurate results. It was found that pose adjustment was necessary in both evaluated models. If the simulation is run without adjustment, odd behaviors could occur such as bone passing through bone or unrealistic joint angles. To prevent this, the initial pose of the model must be adjusted in processing. From sagittal and frontal plane images of the subject, visual approximation was used to modify the model's initial pose by adjusting the coordinates of the DOF in each model. For example, the default initial pose of the Scapulothoracic Joint Model does not seem realistic. The elbow is under the rib cage and the arm is unnaturally straight at the model's side. After adjustment, the arm is in a more natural pose, the elbow is no longer under the rib cage, and there is now space for soft tissue between the arm and torso. The adjustment values used in this study can be found in the table below. Please note that these values should be adjusted between every subject to ensure the most accurate results.
Degree of Freedom | ||
|---|---|---|
Clavicle Prot | N/A | 6 |
Clavicle Elevation | N/A | 1.112 |
Scapula Abduction | N/A | 7.456 |
Scapula Elevation | N/A | 3.732 |
Scapula Upward Rotation | N/A | 5.322 |
Scapula Winging | N/A | -2.080 |
Shoulder Elevation | 25 | 20 |
Elevation Angle | 65 | N/A |
Shoulder Rotation | 15 | N/A |
3. Algorithm and Code
All software related development for this project was done via a C++ script. The code developed in this project was heavily based on an algorithm which converts IMU data from the lower extremities into an OpenSim friendly format. This algorithm was developed by Mazen Al Borno of the NMBL at Stanford. Adjustments were made to the algorithm to map the IMUs to the upper body, correct for model pose, and adjust the orientation of the IMUs. The segment to anatomical body calibration was not modified from the method used on the lower body, although the code is formatted for easy modification if alternative calibration methods were to be explored.
4. Ground Tilt Measurement
Intuition tells us the tilt of the ground IMU should be of great importance to the accuracy of a simulation. This is because the simulated behavior of all other IMUs is directly influenced by the orientation of the 'ground' body. Early simulations made the assumption that the ground IMU was perfectly aligned to the vertical axis, but experimentation shows this is generally false. Additionally, ground IMU tilt is not consistent between subjects or even between trials. For this reason, the tilt of the IMU was measured before each trial and accounted for in the algorithm. The tilt of the IMU was measured using a goniometer.
Results and Discussion
Due to time limitations, only elbow joint angle as a result of elbow flexion was evaluated as part of this project. All relevant data files can be found below.
In the following plots, Ajay's Model refers to the Scapulothoracic Joint Model, and Kate's Model refers to the MoBL-ARMS Dynamic Upper Limb Model.