💻 computer science

Event-Based Motion Estimation via Oriented Distance Fields

This paper introduces Oriented Distance Field Motion Estimation, a low-latency method that achieves sub-pixel accuracy by replacing iterative optimization with a single averaging step over precomputed event distance fields, while simultaneously enabling real-time image deblurring and power-efficient pupil tracking through its generalized application.

Original authors: Lei Sun, Yuqin Ma, Weilun Li, Haoran Liang, Runyi Yang, Kaiwei Wang, Danda Pani Paudel, Luc Van Gool

Published 2026-08-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lei Sun, Yuqin Ma, Weilun Li, Haoran Liang, Runyi Yang, Kaiwei Wang, Danda Pani Paudel, Luc Van Gool

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a camera that does not take pictures in the usual way, capturing a steady stream of snapshots at a fixed speed. Instead, this special sensor acts like a swarm of millions of tiny, independent observers, each watching a single spot of light. They only speak up when that spot changes brightness, reporting the moment with microsecond precision. This technology, known as an event camera, is designed for the world of high speed, where traditional cameras struggle to keep up. When a camera moves quickly, a standard sensor blurs the image because it tries to capture everything that happened during a fraction of a second. An event camera, however, records only the changes, creating a stream of data that directly encodes the path of motion. The challenge for scientists has been how to read this stream fast enough to be useful. Most existing methods try to solve the puzzle of movement by repeatedly guessing and checking, a process that takes time and slows down the very advantage the sensor was built to provide.

A team of researchers has now introduced a new way to read these motion streams that eliminates the need for guessing altogether. They developed a method called Oriented Distance Field Motion Estimation. Instead of running a complex calculation for every single piece of data that arrives, the system first builds a static map based on the edges of the scene it sees. This map acts like a pre-computed guidebook. When a new piece of data arrives, the system simply looks up its position on this map to find the direction and distance the object has moved. It then averages these lookups to determine the overall path. This approach replaces a slow, repetitive search with a single, instant lookup. The result is a system that can track movement with sub-pixel accuracy while operating at a speed that is orders of magnitude faster than previous methods, effectively removing the delay that had previously held back these high-speed sensors.

The researchers tested this new method on both public data and their own recordings, comparing it against the best existing techniques. They found that their system could determine the trajectory of a moving object with an error smaller than a single pixel, matching or beating the accuracy of older methods. More importantly, it did so with a latency that was significantly lower, updating its estimate hundreds of times per second. To prove that this speed and accuracy were useful in the real world, the team applied their findings to two very different tasks. First, they used the motion path to fix blurry photographs. By knowing exactly how the camera moved while a picture was being taken, they could mathematically reverse the blur. They paired their motion data with a small, efficient computer program to restore sharp images, achieving results that were as good as or better than much larger, more complex systems, but using far fewer computing resources.

In a second application, the team used the same motion data to track human eyes. This is a difficult task for standard cameras because the eye moves so fast and the lighting can change rapidly. By using their motion-estimation method to filter out noise and follow the edge of the pupil, they created a tracker that could follow a person's gaze for tens of seconds without losing its place. They built a prototype device for near-eye use, similar to what might be worn in augmented reality glasses. This device consumed less than half the power of a comparable camera system that took regular pictures. The event-based tracker drew only 127.1 milliwatts of power at 60 frames per second, compared to 215.0 milliwatts for the traditional camera, demonstrating that the new method not only works faster but also allows for lighter, more energy-efficient devices.

The success of this approach relies on a few specific conditions. The method works best when the entire scene moves together, such as a camera panning across a landscape, rather than when objects in the scene are moving independently. It also requires that the edges in the scene have enough variety in their angles to provide a clear direction for the movement. If a scene is dominated by a single straight line or if objects are moving in different directions, the accuracy can drop. The researchers acknowledge these limits, noting that their current work focuses on two-dimensional movement, which covers many practical uses like stabilizing a camera or tracking a person's head, but does not yet handle complex three-dimensional rotations. Despite these boundaries, the core discovery stands: by replacing a slow, iterative search with a single, pre-computed lookup, the team has unlocked the full speed potential of event cameras, allowing them to see and react to the world as fast as the light itself changes.

Technical Summary: Event-Based Motion Estimation via Oriented Distance Fields

Problem Statement
Event cameras offer microsecond temporal resolution and high dynamic range, making them ideal for platforms requiring rapid motion recovery, such as mobile robots with gimbal or Pan-Tilt-Zoom (PTZ) systems. However, existing event-based motion estimation methods often negate the sensor's inherent low-latency advantage. Current approaches typically rely on either iterative optimization (e.g., contrast maximization) or discrete hypothesis search (e.g., multi-hypothesis trackers). These methods require repeated evaluation of energy functions or comparison of candidate states for every incoming event, creating a computational bottleneck that prevents the system from keeping pace with fast angular motion.

Methodology: Oriented Distance Field (ODF) Motion Estimation
The paper proposes Oriented Distance Field (ODF) Motion Estimation, a closed-form algorithm that replaces iterative optimization and hypothesis search with a single averaging step. The method operates under the assumption of globally consistent 2-DoF (two degrees of freedom) translation, which is common in scenarios like camera panning or handshake.

The core pipeline consists of three stages:

Adaptive Pattern Generation: Instead of using fixed time windows or event counts, the method builds an edge template by monitoring the "collision rate" of downsampled events. It accumulates events until the fraction of overlapping bins in a new bundle exceeds a threshold, ensuring sufficient spatial coverage without over-densifying the template.
Oriented Distance Field Construction: For a given edge template, the system precomputes a field over the image domain. At every pixel location, this field stores a distance value $D(x)$ and a unit gradient-descent direction $\hat{d}(x)$ . The product $v(x) = D(x)\hat{d}(x)$ represents the translation vector required to align that pixel with the reference edge. This field is computed once per template.
Trajectory Extraction via Averaging: As new events arrive, the algorithm performs a direct lookup of the translation vector $v(x)$ $v (x)$ from the precomputed field. To resolve the aperture problem (ambiguity in motion direction along a single edge), the method averages these vectors over a batch of events. This averaging leverages orientation diversity within the batch to constrain both components of the 2-DoF translation.
- Noise Handling: A parameter-free spatio-temporal contrast filter removes trailing events, and a valid region mask excludes events too far from edges or located in dense edge regions where field values are biased toward zero.
- Efficiency: The per-event cost is reduced to a field lookup and an accumulation step, decoupled from scene complexity or motion magnitude.

Downstream Applications
The authors validate the generality of the ODF estimator through two distinct applications, treating them as tests of the estimator's accuracy and robustness rather than proposing new architectures:

Real-Time Non-Blind Image Deblurring: The estimated trajectory is converted into a continuous, unquantized blur kernel. This kernel is paired with a compact iterative-unfolding network (based on half-quadratic splitting) trained on simulated motion-estimation noise. The network uses an event mask to focus corrections on genuinely blurred regions.
Asynchronous Pupil and Glint Tracking: The same precomputed distance vectors are repurposed as a directional filter for eye tracking. By summing vectors from events near the pupil boundary and filtering out those opposing the estimated motion direction, the system achieves stable tracking of pupil and corneal glints with low power consumption.

Key Contributions

ODF Motion Estimation: A closed-form algorithm for 2-DoF motion estimation that replaces iterative optimization with a single distance-vector average, an adaptive event-count selection strategy, and a parameter-free trail filter.
Generality Validation: Demonstration that the estimator's output is effective across disparate tasks (deblurring and eye tracking) without requiring task-specific retraining of the core estimator.
Efficiency and Accuracy: Extensive evaluation showing that ODF achieves state-of-the-art or competitive accuracy with significantly lower latency (sub-microsecond per event) compared to optimization-based alternatives.

Experimental Results

Motion Estimation: On the Event-Camera Dataset, ODF achieved sub-pixel accuracy (e.g., 0.328 px error on the "slider far" sequence) with update rates up to 1066, significantly outperforming HASTE (467 Hz) and Contrast Maximization (19.5 Hz) in update frequency. While HASTE and CM achieved slightly lower error on specific sequences due to tunable parameters, ODF traded a fraction of a pixel of accuracy for an order-of-magnitude increase in update rate.
Image Deblurring: On the EventAid-B dataset, the proposed method (ODF kernel + lightweight network) achieved 27.23 PSNR and 0.860 SSIM with only 0.6M parameters, outperforming larger blind deblurring networks (e.g., Restormer, NAFNet) and event-assisted baselines like EFNet. On a real-world dataset with ground-truth kernels, the method using the ODF-estimated kernel trailed the ground-truth kernel by less than 1 dB in PSNR, confirming the kernel's high fidelity.
Eye Tracking: The tracker maintained stable pupil and glint tracking for up to 1745 frames (~70 seconds) on public datasets, achieving median IoU scores of 0.85–0.86. On a near-eye prototype, the system processed events in 28 µs (single thread) and reduced total module power draw to 127.1 mW compared to 215.0 mW for a frame-based equivalent at 60 fps.

Significance and Claims
The paper claims that ODF Motion Estimation removes the computational bottleneck that has historically limited the real-time utility of event cameras in fast-motion scenarios. By replacing optimization with a closed-form averaging operation, the method preserves the sensor's low-latency advantage end-to-end. The authors emphasize that the estimator is not tied to a single task; its ability to generate accurate blur kernels and directional filters demonstrates its utility as a general-purpose motion recovery module. The work highlights that high-accuracy motion estimation does not require heavy iterative solvers, enabling efficient, real-time perception on resource-constrained platforms.

Limitations
The authors acknowledge that the method assumes globally consistent 2-DoF translation. Performance degrades in dynamic scenes with independently moving objects, when templates lack orientation diversity (e.g., dominated by a single straight edge), or in sparse event regions. Furthermore, the current applications are limited to 2-DoF translation and do not yet handle general 6-DoF motion or complex affine/projective warps.

More like this