💻 computer science

Quantifying Noise of Dynamic Vision Sensor

This paper introduces a novel technique based on Detrended Fluctuation Analysis (DFA) to quantitatively characterize background activity noise in Dynamic Vision Sensors without ground truth, enabling the derivation of optimal denoising filter parameters.

Original authors: Evgeny V. Votyakov, Alessandro Artusi

Published 2026-03-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Evgeny V. Votyakov, Alessandro Artusi

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 you are trying to listen to a friend whispering a secret in a crowded, noisy room. Your friend's voice is the signal, and the chatter of the crowd is the noise.

This paper is about a special kind of camera called a Dynamic Vision Sensor (DVS). Unlike a normal camera that takes a picture every second (like a video camera), this camera only "sees" changes. If a pixel doesn't change, it stays silent. If something moves or the light changes, it shouts out a tiny "event."

The Problem: The Static in the Signal

The problem with these cameras is that they are a bit "jittery." Even when nothing is moving, they sometimes shout out false alarms. This is called Background Activity (BA) noise. It's like your friend trying to whisper, but the room is so full of random static that you can't tell if a sound is your friend or just a random cough from a stranger.

Usually, engineers try to fix this by setting rules: "If an event happens alone, it's noise. If a bunch of events happen close together in space and time, it's a real object."

But here's the catch: How do you know if you fixed it right?
In a normal photo, you can compare the cleaned photo to the original "perfect" photo. But with these cameras, there is no "perfect" photo to compare against. You don't know what the real scene looked like, so you can't tell if your cleaning method accidentally deleted your friend's voice along with the noise.

The Solution: The "Fingerprints" of Noise

The authors of this paper came up with a clever new way to check the quality of the cleaning without needing a "perfect" photo. They used a statistical tool called Detrended Fluctuation Analysis (DFA).

Here is the analogy:

Real Signal (The Secret): When a real object moves, its events are connected. They follow a pattern, like a line of people walking in a queue. There is order and correlation.
Pure Noise (The Static): Random noise is like people shouting in a crowd with no pattern. It's chaotic and uncorrelated.

The authors realized that if you look at the "noise" left over after you try to clean the data, you can check its "fingerprint."

If the leftover noise still has order (a pattern), it means you didn't clean enough, or you accidentally kept some of the real signal mixed in with the noise.
If the leftover noise is perfectly random (chaotic), you did a great job!

How They Did It (The Experiment)

They took a dataset of events (like a recording of a slot car race) and ran it through a cleaning filter with different settings.

Too gentle: They left the filter settings loose. The "noise" they filtered out still had patterns (like a line of cars). The DFA tool said, "Hey, this noise isn't random enough! You missed some signal."
Too harsh: They cranked the filter up. The noise became perfectly random, but they also started deleting parts of the real car.
The Sweet Spot: They found the "Goldilocks" setting where the leftover noise was as random as possible (statistically speaking, the "scaling exponent" hit a specific number, 0.5). This meant they had successfully separated the signal from the noise without needing to know what the car looked like beforehand.

Why This Matters

This is a big deal because it gives engineers a scientific ruler to measure how well their noise filters are working.

Before: "I think this filter looks good because the video looks clearer." (Subjective guess)
Now: "I know this filter is optimal because the statistical 'randomness' of the leftover noise is exactly where it should be." (Objective fact)

The Takeaway

The paper introduces a way to listen to the "static" in a noisy room and say, "Ah, this static is perfectly random, which means I've successfully isolated the whisper." It allows researchers to tune their cameras to work perfectly in the real world, even when they don't have a "perfect" reference image to compare against.

In short: They found a way to measure the quality of a cleanup job by analyzing the mess that was left behind, rather than trying to find a "before and after" photo that doesn't exist.

1. Problem Statement

Dynamic Vision Sensors (DVS), or event cameras, asynchronously detect per-pixel brightness changes. A critical limitation of DVS technology is Background Activity (BA) noise—spurious events generated by sensor imperfections that are intermixed with useful signal events.

The Challenge: Separating noise from signal is difficult because DVS data is dynamic, and practical scenarios often lack ground truth (a known "clean" reference) to evaluate denoising performance.
Current Limitations: State-of-the-art denoising methods (e.g., spatiotemporal neighborhood checks) rely on the assumption that noise events are isolated. However, assessing the quality of these filters is subjective or computationally expensive. Existing evaluation methods either require synthetic data (which may not reflect reality) or rely on event-wise labeling, which is impractical due to high data rates.
The Gap: There is a lack of an objective, ground-truth-free metric to quantify noise and optimize denoising filter parameters.

2. Methodology

The authors propose a novel approach using Detrended Fluctuation Analysis (DFA), a statistical method typically used to detect long-range correlations in time series, to characterize DVS noise.

A. Data Segmentation

The method involves splitting the raw DVS event stream ( $TS_{full}$ ) into two components using a standard BA filter:

Clean Signal ( $TS_{clean}$ ): Events identified as correlated (signal).
Residual Noise ( $TS_{noise}$ ): Events identified as uncorrelated (noise).

B. Adaptation of DFA to DVS

Standard DFA analyzes a time series $x$ . For DVS, the authors adapt the random variable $x$ to represent the inter-event time intervals ( $\delta t_i = t_i - t_{i-1}$ ).

Cumulative Sum: The cumulative sum of these intervals corresponds exactly to the event timestamps ( $X_i = t_i$ ).
Detrending: The time series is partitioned into segments of length $n$ . A local trend is fitted and subtracted to remove local biases.
Fluctuation Calculation: The root-mean-square fluctuation $F(n)$ is calculated for each segment length.
Scaling Exponent ( $\alpha$ ): By plotting $\log F(n)$ $lo g F (n)$ vs. $\log n$ $lo g n$ , the slope yields the scaling exponent $\alpha$ $α$ .
- $\alpha \approx 0.5$ : Uncorrelated, random noise (ideal for a noise stream).
- $\alpha > 0.5$ : Positive long-range correlations (indicates residual signal structure or non-random patterns).
- $\alpha < 0.5$ : Anti-correlation.

C. Optimization Criterion

The core hypothesis is that an optimal denoising filter will produce a residual noise stream that is as random as possible. Therefore, the optimal filter parameter (specifically the time window $\Delta T$ ) is the one that minimizes the DFA exponent $\alpha$ of the residual noise, driving it toward 0.5.

3. Key Contributions

Novel Metric: Introduction of DFA as an objective criterion for evaluating DVS denoising performance without requiring ground truth data.
Parameter Optimization: A principled method to derive optimal parameters for BA filters (specifically $\Delta T$ ) by statistically analyzing the randomness of the residual noise.
First Application: This is the first known application of DFA to the analysis of event-based camera data.
Validation: Demonstration that the metric works across both simple synthetic-like datasets and complex real-world scenes with injected noise.

4. Results

The authors validated the method on two types of datasets:

Simple Dataset (Slot-car):
- The BA filter was applied with varying time windows ( $\Delta T$ from 1,000 to 16,000 $\mu s$ ).
- Visual vs. Statistical: While visual inspection showed signal traces remaining in the noise at low $\Delta T$ , the DFA metric quantified this.
- Correlation: At $\Delta T = 1000$ , $\alpha = 0.74$ (highly correlated, indicating residual signal). As $\Delta T$ increased to 16,000, $\alpha$ dropped to 0.50, indicating the noise had become truly random.
- Trade-off: The study noted that while maximizing $\Delta T$ improves the metric, it increases computational latency ( $O(N^2)$ for naive implementations). Thus, DFA helps find the "sweet spot" where noise is random without excessive computational cost.
Complex Dataset (PROPHESSEE with Injected Noise):
- Clean event streams were mixed with controlled fractions of BA noise (5%, 10%, 20%).
- As the noise fraction increased, the DFA exponent of the filtered residual noise consistently converged toward 0.5.
- This confirmed that the metric reliably detects effective noise suppression even in complex scenes with moving objects.

5. Significance

Ground-Truth Independence: This work solves a major bottleneck in DVS research by providing a way to tune and evaluate filters in real-world scenarios where ground truth is unavailable.
Objective Standardization: It moves the field away from subjective visual inspection or application-specific metrics (like warped-event contrast) toward a universal statistical standard.
Future Directions: The authors suggest extending this analysis to more advanced filters (e.g., $k$ -noise, $y$ -noise) and exploring the relationship between DFA exponents and Poisson statistics of inter-event intervals.

In summary, the paper establishes that minimizing the DFA scaling exponent ( $\alpha$ ) of the residual noise stream is a robust, objective method for optimizing DVS denoising filters, ensuring that the remaining noise is statistically random and that useful signal has been preserved.