Event-Based Model Fitting

• Event-based model fitting is a technique that robustly estimates parameters and segments structure from asynchronous event data, ensuring temporal and spatial coherence.
• It utilizes adaptive entropy-driven event fusion and deterministic time-slice pairing to generate and prune candidate motion trajectories efficiently.
• A two-stage weighting algorithm combined with robust geometric fitting minimizes noise and outliers, enabling high-precision tracking in neuromorphic and event-driven systems.

Event-based model fitting is a class of methodologies for robustly estimating parameters, segmenting structure, and associating data in domains where observations take the form of asynchronous events. These scenarios arise in neuromorphic vision (event cameras), event-driven sensors, and temporal point process data, where classic frame-based or uniformly sampled approaches are inapplicable or suboptimal. Event-based model fitting emphasizes leveraging asynchrony, local and global structure, and data-adaptive criteria to generate, select, and refine candidate models, all while maintaining computational efficiency and robustness to noise and irrelevant events.

1. Asynchronous Event Data Fusion and Hypothesis Generation

A central challenge in event-based domains, particularly event camera tracking, is the need to aggregate incoming streams of events into sufficiently informative sets for model hypothesis generation without introducing temporal smearing or missing fine-scale dynamics. To this end, the Event Data Association (EDA) framework introduces an adaptive event fusion mechanism based on information entropy computed over an Adaptive Time-Surface with Linear Time Decay (ATSLTD) frame. Formally, for each arriving event $e_k = (t_k, u_k, v_k, p_k)$, the ATSLTD frame $\mathcal{F}$ is updated at the corresponding pixel and decays elsewhere according to the elapsed time, encoding recent event activity in both polarity channels.

When the Non-Zero Grid Entropy $H(\mathcal{F})$, defined as

$H(\mathcal{F}) = -\sum_{i=1}^{B} p_i\log p_i,$

falls within a precomputed Student's-t-based confidence interval $[\alpha, \beta]$ drawn from reference samples in sharp, low-displacement scenarios, the accumulated events over the interval $[t_k, t_k+T]$ are frozen as the new data batch for model fitting. This entropy-based criterion ensures that event batches are neither too sparse (lacking structure) nor too dense (risking merge of multiple structures or blur), preserving spatial-temporal coherence for precise fitting (Chen et al., 2021).

From this fused batch, deterministic model hypothesis generation proceeds by dividing the time interval into $N_s$ slices, pairing events from first and last slices to construct 3D spatio-temporal line (trajectory) hypotheses: $\ell: \mathbf{e}^s=(u_s,v_s,t_s), \mathbf{e}^e=(u_e,v_e,t_e),\quad \mathbf{x}(\lambda)=\left(\begin{array}{c} u_s+\lambda(u_e-u_s)\v_s+\lambda(v_e-v_s)\t_s+\lambda(t_e-t_s) \end{array}\right),\quad \lambda\in[0,1]$ Parallel duplicate hypotheses are pruned by computing the cosine distance, and only distinct representatives with maximal cluster size are retained to form a tractable candidate set.

2. Robust Multi-Stage Model Selection and Weighting

To identify true models among the generated hypotheses, EDA applies a two-stage weighting algorithm (TSW):

• Stage 1 (Temporal Dispersion and Residual Weight): For each line hypothesis, inliers are determined via point-to-line residual thresholding in $\mathbb{R}^3$, and clusters with fewer than three inliers are discarded. A temporal variance weight is assigned:

$$w_j^{(1)} = \frac{1}{N_i\sum_{\mathbf{e}_i \in \mathcal{I}_j} (t_i - \frac{S_t}{2})^2}$$

where $\mathcal{F}$0 is the interval duration.

• Stage 2 (2D Contrast-Based Refinement): The events belonging to each inlier set are projected into a warped image plane, and image contrast is computed:

$\mathcal{F}$1

The final model weight is the product $\mathcal{F}$2.

This two-stage scheme filters out clusters with temporally concentrated inliers (spurious trajectories), low 2D contrast (noise/outliers), or both, leading to high spatial and temporal fidelity in selected models.

Model selection proceeds via adaptive detection of the "elbow" in the sorted weight differences: $\mathcal{F}$3 The local maximum points to the most plausible number of distinct models; the lowest-weight hypotheses up to this point are retained, enabling automatic multi-structure fitting and elimination of both over- and under-segmentation (Chen et al., 2021).

3. Model Fitting, Event Association, and Robust Objective

The final step is a robust assignment of each event to its closest selected model, subject to residual thresholding. Outlier events are rejected, and the collective assignment is equivalent to minimizing a robust aggregate loss: $\mathcal{F}$4 where $\mathcal{F}$5 implements a truncated-quadratic or Huber-style penalty. This geometric multi-structure fitting tightly binds each event to its trajectory, yielding well-defined object motion curves that serve as a basis for down-stream high-precision tasks, such as bounding-box prediction under high-speed, dynamic range, or complex multi-object/camera motions.

4. Algorithmic Outline and System-Level Properties

The above workflow is codified in the following high-level steps:

1. ATSLTD Frame Maintenance: Decay and update per incoming event; compute entropy at each update.
2. Data Freeze on Entropy Interval: When $\mathcal{F}$6, freeze events as a batch, proceed with fitting.
3. Deterministic Hypothesis Generation and Pruning: Time-slice based trajectory construction, cosine-based duplicate removal.
4. Two-Stage Weighting and Contrast Calculation: Residual/inlier computation followed by 2D contrast normalization to obtain final weights.
5. Model Selection via Elbow Criterion: Adaptive cutoff on weight difference to select true models.
6. Event-to-Model Assignment: Robust geometric fitting of all events to selected models.

This pipeline is designed to robustly preserve straight-segment event trajectories, enforce temporal coverage and spatial sharpness, auto-select model numbers, and deliver accurate tracklets by integrating efficient, non-stochastic hypothesis generation with principled, entropy-driven data fusion.

5. Relation to Classical and Neuromorphic Robust Fitting Paradigms

Event-based model fitting marks a departure from classic synchronous robust fitting like RANSAC, in both computational protocol and model hypothesis generation. Whereas RANSAC relies on random minimal-sample subsets and repeated stochastic draws (with associated variance and random coverage risk), EDA’s deterministic time-slice pairing ensures systematic coverage of plausible motion directions and temporal consistency (Chen et al., 2021).

In energy-constrained domains such as neuromorphic hardware, event-based robust fitting can also be mapped to asynchronous, event-driven spiking neural network architectures where spike events only trigger upon data, parameter, or residual changes, providing significant energy savings versus CPU implementations (Nguyen et al., 13 Aug 2025).

The event-based paradigm is also utilized in other domains: e.g., RANSAC-based arrival time event clustering for microseismic event association (Zhu et al., 2017), or in point-process models fit using event sequences and neural architectures with negative evidence for temporal point process parameterization (Gao et al., 2020).

6. Impact, Applications, and Practical Considerations

Event-based model fitting underpins robust, low-latency tracking and segmentation pipelines in event camera-based object tracking, high-speed motion segmentation, asynchronous SLAM, and time-critical interaction systems. The core advances—adaptive, entropy-driven data grouping; deterministic, temporally-aware hypothesis generation; and multi-stage, data-driven model selection—yield superior resilience to sensor noise, low-displacement artifacts, overlapping structures, and adverse operating conditions such as motion blur or lighting extremes (Chen et al., 2021).

Notably, these principles promote clear separation of multiple independently moving objects and accurate per-event associations, supporting high-precision, real-time object tracking and downstream geometric inference in real-world systems where traditional frame-based methods are unsuited.

---

References:

• "Event Data Association via Robust Model Fitting for Event-based Object Tracking" (Chen et al., 2021)
• "Event-driven Robust Fitting on Neuromorphic Hardware" (Nguyen et al., 13 Aug 2025)
• "A multi-channel approach for automatic microseismic event association using RANSAC-based arrival time event clustering(RATEC)" (Zhu et al., 2017)
• "A Multi-Channel Neural Graphical Event Model with Negative Evidence" (Gao et al., 2020)