Page 172 - Biomedical Engineering and Design Handbook Volume 1, Fundamentals

BIOMECHANICS OF HUMAN MOVEMENT 149

to find out there was a problem and the scene needs to be reshot. Researchers are using real-time
information from motion capture to provide patients with visual feedback on how they are performing
a task. Movement patterns can be modified using visual guidance by showing the patient their
pattern (e.g., knee angle) and how it compares to a target pattern. This allows the patient to visualize
how to change their movement pattern to match a desired trajectory. Clinicians are using real-time
motion capture to immerse patients in virtual environments providing a more stimulating exercise
and rehabilitation experience.

Processing Flow for Real-Time Motion Capture. Latency is the key to the “realness” of real-time
motion capture. For our purposes latency is the time difference between the initiation of a movement
and an accurate characterization of the movement as determined by the motion capture system.
A great deal happens between the time a marker is registered by a CCD sensor and the calculation of
three-dimensional kinematics. Not surprisingly, many factors affect latency including the number of
cameras being used, the number of markers that are tracked and of obviously the processing
speed of the CPU. It is helpful to understand the data processing flow from the image sensor to
desktop display when considering latency in the context of real-time motion capture. Although
specific details may vary from system to system, our goal in this section is to provide an overview
of the processing flow and describe how each step adds to the total latency. The steps involved for
a passive marker system can be summarized as follows:
1. Light-emitting diodes on the face of a camera illuminate retroreflective tracking markers in the
camera’s field of view. The diodes are strobed to coincide with the opening of the camera’s
shutter. Light is reflected back off each marker through the camera lens and excites a portion of
the CCD imaging sensor. The exposure time is related to the sampling frequency of the camera
and for some systems the exposure time can be set by the user. For our purposes we will assume
the exposure time is approximately 1 ms at sampling rates typically used when recording human
movement.
2. After exposure, the sensor is scanned and circle fitted to compute the center point of each marker
imaged by the sensor. The time required to do so is approximately 1/maximum frame rate. For
example, a camera that can sample up to 500 fps would require 2 ms to scan the sensor and
circle fit the data.
3. The two-dimensional camera coordinates (u, v) for each marker are packaged and sent to the data
acquisition computer via Ethernet. This is done very efficiently requiring approximately 1 ms for
the number of markers typical of human movement studies (i.e., <40 markers).
4. The data acquisition software must now reconstruct the three-dimensional X, Y, Z coordinates for
each marker and assign the coordinates to a model being tracking. Correct assignment is imper-
ative for accurate model tracking and the time required for this depends on the number of
cameras and markers used. This may take anywhere from 1 to 5 ms. The position and orientation
of the model has now been computed and the data can be sent to another process (i.e., rendered
to screen or sent to third party software).
The processing flow outlined in steps 1 to 4, and the times required for each step are approximate
values. To accurately determine true latency requires access to low-level system architecture.
Qualisys Inc., has reported latency figures for their 6 camera Oqus system tracking 36 markers at
160 fps. The total latency was 6 ms, which corresponds to a delay of 1 video frame. A delay of 1 to
2 video frames is a realistic goal for modern motion capture systems. It is important to note that this
delay is at the level of the motion capture system and does not include additional processing time
associated with rendering to the screen or processing in third party software for visualization. This
can add a significant layer of delay (10 to 30 ms) to the overall latency depending on how the
real-time data are used.

Real-Time Feedback: A Gait Retraining Example. The knee flexes during the initial portion of
stance and helps absorb the impact our body experiences with every step. Normal peak knee flexion
during this time varies from person to person, but is generally between 15° and 25° for healthy adults