Biomedical Engineering Reference
In-Depth Information
advanced facilities are not easily accessible and are complicated to use. Recently,
video analysis techniques have been utilized to analyze ACL injuries by manually
marking these joints and measuring joint angles [
1
,
2
]. Nevertheless, there are few
training programs that use video analysis to train or assess athletes' moving style and
to prevent ACL injuries. Thus, the objective of this paper is to present an effective
approach to estimate these body joints from videos to assess athletes in preventing
ACL injuries. We mainly focus on estimating the lower body joints in lateral and
anterior views separately. We follow the same strategy as in [
3
] by first tracking the
human body parts and then estimating the joint positions defined on human body
parts.
Many methods have been proposed to track human body parts in 2D (e.g., see [
4
]
and references therein). Typically, 2D human body models, such as the stick figure
model [
5
] and the cardboard model [
6
] are used to find the best fit between the
model and image features. In addition, different features such as color [
3
], silhou-
ette [
7
], or blobs [
8
] are commonly used in order to capture human pose variations.
However, much less research has been done to analyze markerless sports
medicine videos for clinical analysis like ACL injury analysis and prevention. It
is not trivial because of the motion blur and severe deformation caused by fast
motion, where the commonly used dynamic models of object motion or appearance
models learned from texture may fail. Another challenge that arose in our applica-
tion is that the background is not stationary. To overcome those difficulties, a
dynamic color model based method is proposed in [
3
] to capture the time-varying
color distributions of jumpers to segment and partition a human body from the
background. However, since only color information is used for segmentation, this
method is sensitive to illumination change and shadows which leads to poor
segmentation. In this work, we propose a new approach to track human body parts
via the consistent propagation of body part labels across frames. In this framework,
video frames are clustered into color blobs, and the relations between those blobs are
represented by a RAG [
9
]. In addition, the geometrical constraints are incorporated
by introducing a loose stick figure to reduce the mismatches in label propagation
process. After tracking the body parts over a sequence, the trajectories of key joints
are estimated and smoothed to reduce errors from online body parts tracking. From
this perspective, our approach suffers less from accumulated errors of online track-
ing and has less computational costs than the classical model fitting methods.
2 Human Body Model
Two
loose
human stick figure models are used to represent the geometrical relations
between human body parts in both anterior and lateral views. Since legs are the
focus of jumping analysis, the stick figure model in an anterior view consists of four
parts, i.e., upper right leg, upper left leg, lower right leg, and lower left leg, which
represent the major axes of their corresponding body parts. The stick figure model
in a lateral view has three components: upper leg, lower leg, and head, based on the
