Inertial Measurement Units (IMUs) are entirely self-contained tracking systems. Because they do not rely on external infrastructure like GPS satellites or cameras, they can operate anywhere: deep indoors, inside elevators, or in dense urban canyons. An IMU directly measures a device's linear acceleration ($a$) and angular velocity ($\omega$).
To find the position of a device from these measurements, a navigation system must perform double integration over time:
$$v(t) = \int a(t) \, dt, \quad p(t) = \int v(t) \, dt$$
While mathematically straightforward, this double integration is highly sensitive to tiny sensor imperfections. Every real-world micro-electromechanical system (MEMS) sensor suffers from deterministic biases and stochastic noise. If an accelerometer has a tiny constant bias $b_a$, integrating it once introduces a linear velocity error ($b_a t$). Integrating it a second time introduces a quadratic position error ($\frac{1}{2}b_a t^2$).
Worse, if the gyroscope has a bias, the system misestimates its orientation relative to gravity. A fraction of the massive gravity vector ($9.81 \text{ m/s}^2$) leaks into the horizontal acceleration channels, causing the position error to grow cubically ($t^3$) over time. Within seconds, an uncorrected IMU will drift by meters; within minutes, the position estimate diverges entirely.
The Ground as an Anchor
To stop this explosive error growth without external positioning infrastructure, we must exploit physical constraints inherent to human and vehicle locomotion.
Consider a pedestrian walking. During the "stance" phase of a gait cycle, the foot plants firmly on the ground. For a brief window of time—known as the midstance phase—the sole of the shoe is completely stationary relative to the earth. Its true physical velocity is exactly zero.
This physical reality forms the basis of the Zero-Velocity Update (ZUPT) technique. As explored by Joonbum Bae in A tele-monitoring system for gait rehabilitation with an inertial measurement unit and a shoe-type ground reaction force sensor, identifying these precise gait phases allows us to treat the ground as an intermittent spatial anchor. This periodic correction prevents small sensor errors from accumulating into massive positioning inaccuracies over time, making tracking highly accurate even when individuals intentionally vary their walking patterns, as validated by Lin Zhou et al. in Validation of an IMU Gait Analysis Algorithm for Gait Monitoring in Daily Life Situations.
This concept extends far beyond foot-mounted sensors:
- Vertical Tracking: Wonsang Song et al. apply ZUPT to vertical indoor tracking in Improving the Vertical Accuracy of Indoor Positioning for Emergency Communication, resetting vertical velocity to zero whenever an elevator stops at a floor.
- Stationary Devices: Otto Seiskari et al. leverage zero-motion periods in HybVIO: Pushing the Limits of Real-time Visual-inertial Odometry to prevent filter collapse and drift when a tracking device remains completely stationary.
- Hybrid Systems: ZUPT is widely used to augment other positioning systems; for instance, Zun Niu et al. integrate ZUPT with smartphone RTK GPS in RTK with the Assistance of an IMU-Based Pedestrian Navigation Algorithm for Smartphones to maintain continuous tracking in challenging urban canyons where GPS signals drop out.
Why ZUPT Works: The Mechanics of the Error-State Filter
It is tempting to think of ZUPT as simply a "hard reset" that forces the velocity vector back to zero. While a hard reset prevents velocity from drifting further, it does nothing to fix the underlying causes of the drift: the sensor biases and orientation errors.
Instead, modern systems implement ZUPT within an Error-State Kalman Filter (ESKF), as detailed by Boqi Wu et al. in Indoor Pedestrian Navigation Research Based on Zero Velocity Correction and Sliding Window.
In an ESKF, the filter tracks the errors in the system's state (position error, velocity error, orientation error, and sensor biases) rather than the total states themselves. The IMU continuously integrates acceleration and angular velocity in an open loop to provide a high-frequency, low-latency trajectory.
When the system detects that the sensor is stationary, it creates a "pseudo-measurement": it tells the filter that the true velocity is exactly zero ($v_{\text{true}} = 0$).
IMU Readings ----> [ Open-Loop Double Integration ] ----> Trajectory Estimate
|
Is it stationary? (ZUPT)
|
v
[ Velocity Error = Estimate - 0 ]
|
v
[ Error-State Kalman Filter ]
|
+-------------------------+-------------------------+
| | |
v v v
Correct Position Correct Orientation Estimate Sensor Biases
The difference between the IMU's currently integrated velocity ($\hat{v}$) and the true velocity ($0$) is the velocity error. Because the filter maintains a covariance matrix that models how velocity errors correlate with orientation errors and sensor biases over time, this single piece of information allows it to mathematically back-propagate corrections.
The filter estimates and subtracts the accelerometer bias, corrects the tilt (orientation) of the device, and bounds the position error growth. By converting a continuous integration process into a series of short, bounded intervals, ZUPT transforms exponential drift into a highly manageable, linear accumulation. This allows for the precise calculation of clinical metrics like stride length and foot clearance for tracking neurodegenerative conditions, as demonstrated by Johannes C. M. Schlachetzki et al. in Wearable sensors objectively measure gait parameters in Parkinson’s disease and Francesca Salis et al. in A multi-sensor wearable system for the assessment of diseased gait in real-world conditions.
The Challenge of Stationary Detection
The elegance of ZUPT depends entirely on the system's ability to accurately detect when it is truly stationary. If the filter mistakenly applies a zero-velocity update while the sensor is still moving, it will aggressively strip away actual motion, destroying the trajectory.
To prevent this, researchers use sophisticated statistical tests. Boqi Wu et al. utilize a Generalized Likelihood Ratio Test (GLRT) over a sliding window of inertial data to evaluate the probability that the foot is stationary. In scenarios with highly dynamic movements like running or climbing stairs, Mingrong Ren et al. demonstrate in A Novel Pedestrian Navigation Algorithm for a Foot-Mounted Inertial-Sensor-Based System that a finite state machine tracking speed variations across the gait cycle is required to reliably isolate stationary intervals.
Physical hardware design also dictates detection accuracy. Arne Küderle et al. discovered in The placement of foot-mounted IMU sensors does affect the accuracy of spatial parameters during regular walking that embedding the IMU inside the shoe sole outperforms clipping it externally. The shoe's material naturally damps high-frequency impact vibrations, providing a cleaner signal for the stationary detection algorithms.
For vehicles, Abhishek Gupta et al. show in Road Grade Estimation Using Crowd-Sourced Smartphone Data that monitoring for periods of low acceleration and low jerk (the rate of change of acceleration) provides stable snapshots to isolate the gravity vector and reset orientation drift.
Expanding the Constraint Horizon
While a perfect zero-velocity state is ideal, the broader insight of ZUPT has been expanded to include other kinematic and environmental constraints:
Non-Holonomic Constraints (NHC)
Pedestrians and wheeled vehicles cannot instantly slide sideways or fly vertically through the air. As shown by Chunyang Yu et al. in Map-Based Indoor Pedestrian Navigation Using an Auxiliary Particle Filter, enforcing these lateral and vertical velocity constraints within a Kalman filter further constrains drift even when the system is in motion.
Contact and Collision Constraints
What if a robot never stops moving? Thomas Lew et al. introduce Contact Inertial Odometry in Contact Inertial Odometry: Collisions are your Friends. Instead of waiting for a prolonged pause, they treat physical impacts as instantaneous velocity constraints. By assuming the velocity perpendicular to a contact surface drops to zero at the exact millisecond of impact, the system bounds velocity drift using the physics of collisions.
Bidirectional Smoothing
In mobile gait analysis where real-time execution isn't mandatory, we can look both forward and backward in time. Julius Hannink et al. demonstrate in Benchmarking Foot Trajectory Estimation Methods for Mobile Gait Analysis that fusing forward integration with time-reversed integration perfectly satisfies the zero-velocity boundary conditions at both the start and end of a stride, drastically reducing errors in foot clearance calculations.
Context-Aware Corrections
Advanced systems also apply specialized constraints based on detected motion states. For instance, Fangmin Li et al. show in 3D Tracking via Shoe Sensing that analyzing the "distance error gradient" during ZUPT periods can specifically isolate and eliminate vertical tracking errors during stair climbing. To handle horizontal maneuvers, Haonan Jia et al. introduce a gait-phase-aware turning correction module in A Novel Integrated IMU-UWB Framework for Walking Trajectory Estimation in Non-Line-of-Sight Scenarios Involving Turning Gait to dynamically adjust constraints based on turning intensity, preventing overcorrection during complex maneuvers.
Key Foundational Work
Improving the Vertical Accuracy of Indoor Positioning for Emergency Communication (2012) — contributed 3 concept elements to this direction.
A tele-monitoring system for gait rehabilitation with an inertial measurement unit and a shoe-type ground reaction force sensor (2013) — contributed 2 concept elements to this direction.