Inertial Measurement Units (IMUs) are the workhorses of modern motion tracking. Found in everything from smartphones to high-end motion capture suits, these sensors—comprising accelerometers and gyroscopes—offer high-frequency data without requiring a line-of-sight. However, they suffer from a fundamental mathematical flaw: cumulative drift.
Because IMUs measure changes in motion (acceleration and angular velocity) rather than absolute position, a tracking system must "dead reckon" its current state by integrating these measurements over time. As Changhao Chen and colleagues note in OxIOD: The Dataset for Deep Inertial Odometry, even tiny amounts of sensor noise or bias are integrated twice to calculate position. This causes errors to grow quadratically, leading to a "drift" where the digital representation of a person or robot rapidly diverges from its physical location.
The Complementary Strengths of Fusion
To solve the drift problem, researchers have converged on a shared strategy: sensor fusion. The intuition is to pair the "accurate but drifting" local motion of an IMU with a "drift-free but noisy" global reference.
In Human POSEitioning System (HPS), Vladimir Guzov and Aymen Mir demonstrate this by fusing head-mounted cameras with body-worn IMUs. While the camera provides absolute spatial anchoring (preventing the subject from "drifting" through walls), the IMUs fill in the high-frequency details of limb movement that the camera might miss due to occlusion or motion blur.
This "Rosetta Stone" approach, as described by Saeed Ghorbani in the MoVi dataset, allows systems to maintain long-term accuracy in naturalistic settings. By using external references—whether they are multi-view images, LiDAR, or even magnetic fields—researchers can periodically "reset" the IMU's accumulated error.
Anchoring with Vision and LiDAR
Vision is the most common anchor. In Recovering Accurate 3D Human Pose in the Wild Using IMUs and a Moving Camera, the authors use a single handheld camera to provide a global coordinate frame for a set of wearable IMUs. This corrects the "heading drift" that typically causes a reconstructed skeleton to rotate erroneously over time.
However, vision fails in low light or smoke. In these "sensor-hostile" environments, researchers like Chengfeng Zhao have turned to LiDAR. In LiDAR-aid Inertial Poser (LIP), the system uses the global geometric strength of LiDAR point clouds to constrain the local dynamic precision of sparse IMUs. Similarly, Yudi Dai and the team behind HSC4D use hip-mounted LiDAR to create "digital twins" of humans interacting with massive 5,000 $m^2$ spaces, a feat impossible with drifting IMUs alone.
Neural Approaches to Dead Reckoning
While traditional filters (like the Kalman filter) have long been used for fusion, recent work explores whether deep learning can model and subtract the drift itself.
Sijie JI and Chenshu Wu introduced NeurIT, which uses a Time-Frequency Block-recurrent Transformer to learn the complex, non-linear patterns of sensor bias. By training on vast amounts of data where the ground truth is known, these models learn to "predict" the drift before it happens, achieving sub-meter errors over long distances without constant external fixes.
Even in extreme environments where GPS and vision are both unavailable—such as underwater or deep underground—the principle remains the same. Xueyun Wang and Bill Moran have explored using Quantum Diamond Magnetometry to provide a "position fix" based on Earth's magnetic map, effectively halting the inherent drift of inertial navigation systems.
Summary of the Intuition
The research direction across these 314 papers clarifies a single reality: Inertial sensors are excellent at telling you how you moved in the last millisecond, but they are blind to where you are in the world. By anchoring these high-speed local measurements to a stable global reference—be it a camera frame, a LiDAR scan, or a magnetic map—we can build tracking systems that are both responsive and enduringly accurate.