Receding-Horizon Maximum-Likelihood Estimation of Neural-ODE Dynamics and Thresholds from Event Cameras

Kazumune Hashimoto, Kazunobu Serizawa, Masako Kishida

Abstract: Event cameras emit asynchronous brightness-change events where each pixel triggers an event when the last event exceeds a threshold, yielding a history-dependent measurement model. We address online maximum-likelihood identification of continuous-time dynamics from such streams. The latent state follows a Neural ODE and is mapped to predicted log-intensity through a differentiable state-to-image model. We model events with a history-dependent marked point process whose conditional intensity is a smooth surrogate of contrast-threshold triggering, treating the contrast threshold as an unknown parameter. The resulting log-likelihood consists of an event term and a compensator integral. We propose a receding-horizon estimator that performs a few gradient steps per update on a receding horizon window. For streaming evaluation, we store two scalars per pixel (last-event time and estimated log-intensity at that time) and approximate the compensator via Monte Carlo pixel subsampling. Synthetic experiments demonstrate joint recovery of dynamics parameters and the contrast threshold, and characterize accuracy–latency trade-offs with respect to the window length.

Introduction

Event cameras, such as the Dynamic Vision Sensor (DVS), represent a paradigm shift in computer vision. Unlike traditional cameras that capture frames at fixed intervals, event cameras report asynchronous pixel-level brightness changes. This results in microsecond-level temporal resolution and high dynamic range. However, the data is irregular and history-dependent: a pixel only fires if the log-intensity change since its last event exceeds a specific contrast threshold. These thresholds are typically unknown and can vary across the pixel array.

Most existing approaches simplify event data by aggregating it into fixed-time "voxel grids," which discards precise timing information. This paper addresses the challenge of performing online maximum-likelihood estimation (MLE) of continuous-time dynamics directly from raw event streams while simultaneously learning the sensor's internal thresholds.

Methodology

The proposed architecture bridges the gap between continuous latent dynamics and discrete, asynchronous events through three core components:

  1. Neural ODE Dynamics: The system's latent state (e.g., the position of a moving object) is modeled using a Neural Ordinary Differential Equation. This allows the state to evolve continuously, naturally matching the asynchronous nature of the sensor.
  2. Differentiable Observation Model: A renderer maps the latent state to a predicted log-intensity image. To account for history-dependence, the model tracks the time and log-intensity of the last event for every pixel.
  3. Smooth Surrogate Likelihood: The "hard" event-triggering threshold is replaced with a smooth surrogate function within a marked temporal point process. This makes the entire pipeline differentiable, enabling gradient-based optimization of both dynamics parameters ($\vartheta$) and threshold parameters ($\psi$).
The system architecture integrates online data collection from an event camera with a Neural ODE dynamics model and a receding-horizon MLE framework.

To enable real-time operation, the authors employ Receding-Horizon Estimation (RHE). Instead of optimizing over the entire history, the estimator performs gradient steps on a fixed-lag window of recent events. The computational cost of the "compensator" term (representing the probability of no event occurring) is managed via Monte Carlo pixel subsampling.

Experiments and Results

The framework was evaluated using synthetic event-camera sequences of a moving Gaussian blob. The blob followed a "stable focus" trajectory (spiraling motion), and the sensor was simulated with a spatially non-uniform threshold map.

Joint Parameter Recovery

The estimator successfully recovered both the complex dynamics parameters and the spatial structure of the per-pixel threshold map. As shown in the convergence plots, the estimates for the damping coefficient ($\hat{\alpha}$) and angular velocity ($\hat{\omega}$) converged toward their respective ground truth values over successive episodes.

Convergence of estimated dynamics parameters $\hat{\alpha}$ and $\hat{\omega}$ toward ground truth (GT) values over 30+ episodes of online updates.

Performance Trade-offs

A "horizon ablation" study revealed that longer receding horizons ($H$) significantly improve the estimation of sensitive parameters like angular velocity but increase computational latency. In a standard GPU environment, the mean update time was approximately 300–330ms, which is faster than the 400ms update interval, confirming the potential for real-time deployment.

Conclusion

This work demonstrates that integrating Neural ODEs with a probabilistic point-process model allows for the joint estimation of object dynamics and sensor characteristics from raw event data. By treating contrast thresholds as learnable parameters and utilizing a receding-horizon approach, the method achieves accurate online estimation. Future research will focus on scaling to higher-resolution sensors and more complex, non-synthetic scene structures.

Summary updated

Abstract

Event cameras emit asynchronous brightness-change events where each pixel triggers an event when the last event exceeds a threshold, yielding a history-dependent measurement model. We address online maximum-likelihood identification of continuous-time dynamics from such streams. The latent state follows a Neural ODE and is mapped to predicted log-intensity through a differentiable state-to-image model. We model events with a history-dependent marked point process whose conditional intensity is a smooth surrogate of contrast-threshold triggering, treating the contrast threshold as an unknown parameter. The resulting log-likelihood consists of an event term and a compensator integral. We propose a receding-horizon estimator that performs a few gradient steps per update on a receding horizon window. For streaming evaluation, we store two scalars per pixel (last-event time and estimated log-intensity at that time) and approximate the compensator via Monte Carlo pixel subsampling. Synthetic experiments demonstrate joint recovery of dynamics parameters and the contrast threshold, and characterize accuracy--latency trade-offs with respect to the window length.
BibTeX
@misc{hashimoto2026receding,
  author = {Kazumune Hashimoto and Kazunobu Serizawa and Masako Kishida},
  title = {{Receding-Horizon Maximum-Likelihood Estimation of Neural-ODE Dynamics and Thresholds from Event Cameras}},
  year = {2026},
  eprint = {2603.05011},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.05011},
}
Figure