Most motion capture systems share a convenient lie: they treat the human body as a collection of rigid sticks connected by perfect hinges. In reality, humans are composed of shifting muscle, elastic skin, and fluctuating fat. When we attach a sensor to a person's thigh, we aren't measuring the femur; we are measuring a point on a "bag of water" that wobbles, slides, and compresses independently of the bone.
This discrepancy is known as the Soft Tissue Artifact (STA). Across biomechanics and computer vision, researchers are moving beyond the rigid-body assumption to model the complex, non-rigid relationship between the surface we can see and the skeleton we want to track.
The Problem of the Shifting Surface
The fundamental challenge in human movement tracking is that sensors—whether optical markers, IMUs, or smart textiles—are decoupled from the underlying anatomy. As Huosheng Hu notes in CSM-420, high sensor precision does not guarantee high accuracy if the skin slides or wobbles during activity.
This "noise" isn't random; it is a systematic physical deformation. For example, when a sprinter reaches maximum velocity, sensors on the upper back may fluctuate in ways that do not reflect the athlete's actual center of mass, as explored by Tiago Guedes Russomanno in Assessment of Sprint Parameters. Even in clinical settings, such as swallowing assessments, Hiroaki Hashimoto found that surface markers on the neck fail to perfectly mirror the internal displacement of the larynx in Non-invasive quantification of human swallowing.
Seeing Through the Tissue
To solve the STA problem, researchers first had to quantify exactly how much the skin moves relative to the bone. This requires "gold standard" data where the internal skeleton is visible.
One approach uses medical imaging to create high-fidelity ground truths. Georg Stillfried and Patrick van der Smagt used MRI to "see through" soft tissue, creating a 24-degree-of-freedom hand model that maps bone movement with precision impossible for surface markers (Human hand kinematics based on MRI imaging). Similarly, Dalton D. Moore validated markerless pose estimation against 3D X-ray radiography (XROMM), explicitly modeling the displacement of skin and fur to distinguish between algorithm errors and physical tissue movement (Validating marker-less pose estimation with 3D x-ray radiography).
Perhaps the most direct validation comes from Reference in-vitro dataset for inertial-sensor-to-bone alignment, where sensors were rigidly pinned directly into the bones of cadaveric models. This eliminates STA entirely, providing a baseline to measure exactly how much "noise" skin-mounted devices actually introduce.
Compensating for the Wobble
Once the nature of soft tissue movement is understood, the goal shifts to compensation. Modern research employs three primary strategies to bridge the skin-to-bone gap:
1. Physics-Based Optimization Instead of treating each frame in isolation, researchers like Marlies Nitschke and Anne D. Koelewijn use musculoskeletal simulations to constrain movement. By tracking raw marker data through a physics engine, they ensure that the reconstructed motion is dynamically consistent, effectively filtering out non-physical "jiggles" that violate skeletal constraints (Change the direction).
2. Learned Secondary Motion When sensors are attached to loose clothing, the problem intensifies. Chengxu Zuo and colleagues developed the "Loose Inertial Poser," which uses a specialized AutoEncoder to model "secondary motion"—the way a jacket sleeve slides and twists against the arm. This allows for accurate tracking even when the sensor is not tightly strapped to the body (Loose Inertial Poser).
3. Synthetic Augmentation To train more robust AI models, researchers are now simulating these artifacts in virtual environments. Dominik Hollidt introduced EgoSim, which uses a spring-damper mounting system to replicate the realistic lag and jiggle of body-worn cameras (EgoSim). Similarly, the MODA technique simulates the natural "drift" and displacement of IMUs caused by muscle contraction, forcing models to learn features that are invariant to sensor shifting (MODA: Motion-Drift Augmentation).
Toward Automated Biomechanics
The ultimate aim of this research is to move high-precision biomechanical analysis out of the lab. Platforms like AddBiomechanics, developed by Keenon Werling and others, automate the scaling and optimization process, reducing the time to generate musculoskeletal simulations from hours to minutes. By mathematically accounting for the non-rigid reality of the human body, these systems allow us to finally "see" the skeleton through the noise of the skin.