IMU Preintegration

A technique that pre-integrates a large batch of IMU readings between two frames into a single relative-motion constraint.

An IMU (inertial measurement unit) outputs angular velocity and acceleration at several hundred hertz, while camera or lidar keyframes come at only ten to a few tens of hertz. An optimization-based visual-inertial odometry system needs to add an IMU constraint between neighboring keyframes, but directly integrating depends on the pose and velocity at the starting point — every time optimization changes that starting point, the integration would need to be redone, which is slow. Preintegration instead integrates this stretch of IMU data, expressed in the body frame at the starting point, into relative rotation, velocity, and position increments that don’t depend on the global pose, so it only needs to be computed once; when the estimated bias (the IMU’s systematic offset) changes, a first-order approximation corrects it without recomputing from scratch. Forster and colleagues gave a complete derivation on the rotation manifold SO(3) in IEEE T-RO in 2016, and this approach has become a standard component in visual/lidar-inertial odometry systems like VINS-Mono and LIO-SAM, and in factor-graph libraries like GTSAM.

ExampleVINS-Mono, in its tightly coupled nonlinear optimization, preintegrates the IMU data between neighboring keyframes into a single constraint, solved jointly with visual feature observations to get the camera trajectory.

Also called

IMU Preintegration Factor, Inertial Preintegration

Related

Inertial Measurement Unit · Visual-Inertial Odometry · Factor Graph Optimization · VINS-Mono / VINS-Fusion · Tightly-Coupled vs. Loosely-Coupled Fusion · LiDAR-Inertial Odometry

Sources

arXiv 1512.02363: On-Manifold Preintegration for Real-Time Visual-Inertial Odometry（IEEE T-RO 2016）
arXiv 1708.03852: VINS-Mono