Event Stream Filtering via Probability Flux Estimation

TL;DR

Event streams from neuromorphic cameras mix discrete state information at event times with continuous process information between events, yet traditional filters discard the latter. EDFilter provides a probability‑flux–theoretic framework that estimates threshold‑crossing fluxes at the contrast boundaries using nonparametric, kernel‑based methods and reconstructs a continuous event density flow in real time via an $O(1)$ recursive solver; motion priors further regularize the density with a 4‑directional basis. The Rotary Event Dataset (RED) is introduced to supply microsecond ground truth irradiance references for rigorous evaluation. Across denoising, tracking, SLAM, and video reconstruction tasks, EDFilter demonstrates superior temporal fidelity, physically interpretable denoising, and real‑time performance, highlighting probability‑flux modeling as a viable physics‑informed paradigm for event‑based vision.

Abstract

Event cameras asynchronously capture brightness changes with microsecond latency, offering exceptional temporal precision but suffering from severe noise and signal inconsistencies. Unlike conventional signals, events carry state information through polarities and process information through inter-event time intervals. However, existing event filters often ignore the latter, producing outputs that are sparser than the raw input and limiting the reconstruction of continuous irradiance dynamics. We propose the Event Density Flow Filter (EDFilter), a framework that models event generation as threshold-crossing probability fluxes arising from the stochastic diffusion of irradiance trajectories. EDFilter performs nonparametric, kernel-based estimation of probability flux and reconstructs the continuous event density flow using an O(1) recursive solver, enabling real-time processing. The Rotary Event Dataset (RED), featuring microsecond-resolution ground-truth irradiance flow under controlled illumination is also presented for event quality evaluation. Experiments demonstrate that EDFilter achieves high-fidelity, physically interpretable event denoising and motion reconstruction.

Figures (13)

• Figure 1: Interpretation of the information encoded in an event stream. Each event provides (1) state information through the discrete irradiance jump at the event time, and (2) process information by constraining the latent irradiance trajectory between events through the contrast-bound inequality.
• Figure 2: Stochastic analysis of event generation for a single pixel.
• Figure 3: The proposed EDFilter. The density-prediction module sequentially applies KDE to the input events and selects the optimal kernel by maximizing the event-observation likelihood, producing a predicted density sample. The density-update module then spatially fuses these density samples using a motion-aware, sparsity-preserving local filter to obtain refined density estimates. Finally, the event-sampling module interpolates the continuous event-density flow via zero-order hold and resamples the filtered events for output. This output also triggers the density-prediction module in an application-dependent manner, mitigating the impact of abnormal integration.
• Figure 4: System setup of the Rotary Event Dataset benchmark.
• Figure 5: Visualization of filtered events as 2D event frames (top 3 rows) and 3D point clouds (bottom row). Sequences come from the proposed RED dataset (rows 1,4), the E-MLB dataset ding2023mlb (row 2) and the ECD dataset mueggler2017event (row 3).

Theorems & Definitions (21)

• Proposition 6.1
• proof
• Proposition 7.1
• proof
• Proposition 7.2
• proof
• Proposition 7.3
• proof
• Proposition 7.4
• proof