In a laboratory setting, capturing human movement is a solved problem. By instrumenting a subject with dozens of retroreflective markers or a high-density suit of inertial measurement units (IMUs), researchers can reconstruct every nuance of a gait cycle or a golf swing. However, as Ricotti et al. point out, these high-density configurations present a steep trade-off: the more sensors a system requires, the less likely a patient is to wear it in daily life.
The emerging research direction of sparse sensing aims to break this trade-off. The goal is to reconstruct full-body kinematics, joint loads, or complex activities using a minimalist set of sensors—sometimes as few as one or two. This works not because the missing data is unnecessary, but because human movement is governed by deep structural redundancies and physical constraints that allow a few well-placed observations to represent the whole.
The Geometry of Constraint
The primary reason sparse sensing works is that the human body is not a collection of independent parts, but a kinematically linked chain. When you move your knee, your hip and ankle must follow specific trajectories dictated by bone lengths and joint limits.
Zhang et al. demonstrated this by showing that sensors localized to a single joint—in their case, an "intelligent knee sleeve"—can accurately estimate the orientation of adjacent joints. By leveraging these inherent kinematic constraints, machine learning models can "infer" the position of the foot or hip simply by observing the state of the knee.
This principle extends to the entire body. von Marcard et al. showed that just six IMUs are sufficient to reconstruct accurate 3D human poses in unconstrained outdoor environments. The system doesn't need a sensor on every bone because the skeletal structure itself acts as a physical prior, filtering out impossible configurations and narrowing the search space for the underlying motion.
The Informational Value of the Center of Mass
If you can only place one sensor on a human, where should it go? A recurring finding across the literature is the disproportionate value of the body's center of mass.
As early as 2005, Sung et al. found that high accuracy in activity recognition could be achieved using a single sensor placed near the trunk. This is because the torso acts as a low-pass filter for the movements of the limbs; it captures the global intent of the body. Similarly, Davis et al. utilized a single sacral-mounted IMU to capture center-of-mass dynamics, which provided enough signal information to estimate vertical ground reaction forces during running.
This "minimalist" approach is even effective for complex tasks like load monitoring. Benocci et al. demonstrated that a single chest-mounted sensor could detect not just if a person was carrying a load, but how they were carrying it, by analyzing subtle alterations in gait and trunk rotation.
Strategic Placement and Redundancy
Sparse sensing is not just about using fewer sensors; it is about using the right sensors. Because joint trajectories are highly redundant—a concept explored by Christian et al.—many data points in a full-body suit provide overlapping information.
The challenge, then, is an optimization problem: finding the "sweet spots" on the body where the signal-to-noise ratio for a specific movement is highest. Xu et al. addressed this by using a graph-based framework to optimize the placement of Forcemyography (FMG) sensors. They found that just four strategically placed sensors could match the performance of a 16-sensor array, proving that strategic localization can compensate for a drastic reduction in hardware.
This redundancy is also present in physiological signals. Sui et al. found that standing quality in patients with spinal cord injuries could be classified with a reduced sensor set because multi-channel muscle activity (EMG) is inherently coupled. Similarly, Isaias et al. achieved optimal gait event prediction in Parkinson’s patients using only two EMG probes, bypassing the need for full-leg instrumentation.
Filling the Gaps with Temporal Priors
When sensors are sparse, the resulting data can be "jittery" or ambiguous. To solve this, researchers use hybrid models that combine the pattern-recognition power of neural networks with the temporal consistency of physics-based filters.
Wouda et al. illustrated this by passing neural network predictions through an error-state Kalman Filter. While the neural network might suggest a physically impossible jump between frames due to sparse input, the Kalman Filter enforces temporal smoothness, ensuring the reconstructed motion follows the laws of physics.
By combining these three pillars—kinematic constraints, strategic placement at the center of mass, and temporal filtering—researchers are moving toward a future where "lab-grade" biomechanical analysis can be performed using the devices already in our pockets or the fabric of our clothes.
Key Foundational Work
Wearable feedback systems for rehabilitation (2005) — contributed 1 concept elements to this direction.
Using sensor orientation information for computational head stabilisation in 3d electromagnetic articulography (EMA) (2009) — contributed 1 concept elements to this direction.