In traditional computer vision and robotics, motion is typically treated as a sequence of discrete snapshots. Whether it is a series of camera poses in a SLAM pipeline or a sequence of frames in a video, we assume the world stands still for an instant, moves, and stands still again. However, real-world sensors—from rolling-shutter CMOS cameras to scanning LiDARs and asynchronous event streams—do not capture the world in discrete instants. They integrate information over time.
Continuous-time trajectory modeling replaces the discrete "pose-to-pose" paradigm with a smooth, parametric function of time, typically using B-splines or Bézier curves. By representing motion as a continuous curve in $SE(3)$ space, researchers can query the exact state of a system at any arbitrary timestamp, enabling the correction of sensor distortions and the extraction of information from motion blur.
The Spline as a Physical Backbone
The core mechanism of this research direction is the use of mathematical splines to represent the path of an object or camera. Unlike a sequence of independent matrices, a spline provides inherent temporal smoothness and physical plausibility.
Droeschel and Kerl et al. were early proponents of using cumulative cubic B-splines in the $SE(3)$ Lie algebra. This representation allows for the interpolation of poses at any timestamp while ensuring that velocity and acceleration remain continuous. This is not just a mathematical convenience; it acts as a powerful regularizer. As Kerl et al. demonstrate, this continuous formulation can reduce translational drift by over 90% in environments where discrete-time methods struggle with noise or rapid movement.
In the context of modern dynamic rendering, Lee et al. and Park et al. show that explicit splines (like Cubic Hermite Splines) can replace heavy neural networks (MLPs) for modeling 4D trajectories. This shift maintains the high rendering quality of 3D Gaussian Splatting while significantly reducing computational latency, as the motion is defined by learnable control points rather than expensive network queries.
Correcting the Jello Effect
Most modern cameras use a rolling shutter, where image rows are exposed sequentially from top to bottom. If the camera moves during this process, the resulting image suffers from "jello" distortion—straight lines appear curved because the top of the image was captured at a different time than the bottom.
Continuous-time modeling solves this by assigning a unique timestamp to every row of an image. Li et al. (USB-NeRF) and Liao et al. demonstrate that by querying a B-spline trajectory for each row's specific timestamp, one can mathematically "unroll" the shutter and recover the true 3D geometry. Similarly, in LiDAR scanning, where a laser rotates to capture a point cloud, the sensor is constantly moving. Park et al. and Quenzel et al. use continuous trajectories to de-distort these scans, ensuring that every laser point is projected from its exact position in space at the microsecond it was fired.
Turning Blur into Information
While motion blur is often viewed as noise, it is actually a temporal record of a trajectory integrated over the camera's exposure window. If we know the continuous path the camera took while the shutter was open, we can synthesize what the blurred image should look like.
Zhao et al. (BAD-Gaussians) and Chugunov et al. leverage this by jointly optimizing a continuous trajectory and a 3D scene representation. By "re-blurring" the 3D model along the estimated spline path to match the blurry input, they can recover sharp 3D scenes from handheld, shaky photography. Zhao et al. even extend this to human pose, showing that the "smear" of a limb in a single blurry frame contains enough information to reconstruct the continuous arc of a movement.
Bridging Asynchronous Sensors
Continuous-time modeling is particularly vital for event cameras, which do not capture frames at all, but rather a stream of asynchronous brightness changes. Hamann et al. and Liao et al. show that by modeling motion as a continuous Bézier or B-spline curve, one can integrate these asynchronous events over time to estimate dense optical flow or camera egomotion. This approach bypasses the need to artificially group events into "frames," allowing the system to track agile, high-speed maneuvers that would be impossible for traditional frame-based cameras.
Smoothness and Kinodynamic Constraints
Beyond reconstruction, continuous-time representations are essential for robotics. Kicki et al. and Xu et al. use B-splines to ensure that robot trajectories are not just geometrically feasible, but kinodynamically smooth. By decoupling the spatial path from the temporal scaling, these methods allow for the analytical derivation of velocity and acceleration, ensuring a robot never attempts a move that would violate its physical limits.
Whether correcting a "jello" video or planning a robot's flight, the shift from discrete steps to continuous curves allows algorithms to respect the underlying physics of time and motion. As Mirvakhabova argues in the context of Neural ODEs, integrating along a continuous path prevents the overshooting errors inherent in discrete updates, leading to more stable and accurate motion estimation across the board.