Noise Filtering Benchmark for Neuromorphic Satellites Observations

Abstract

Event cameras capture sparse, asynchronous brightness changes which offer high temporal resolution, high dynamic range, low power consumption, and sparse data output. These advantages make them ideal for Space Situational Awareness, particularly in detecting resident space objects moving within a telescope's field of view. However, the output from event cameras often includes substantial background activity noise, which is known to be more prevalent in low-light conditions. This noise can overwhelm the sparse events generated by satellite signals, making detection and tracking more challenging. Existing noise-filtering algorithms struggle in these scenarios because they are typically designed for denser scenes, where losing some signal is acceptable. This limitation hinders the application of event cameras in complex, real-world environments where signals are extremely sparse. In this paper, we propose new event-driven noise-filtering algorithms specifically designed for very sparse scenes. We categorise the algorithms into logical-based and learning-based approaches and benchmark their performance against 11 state-of-the-art noise-filtering algorithms, evaluating how effectively they remove noise and hot pixels while preserving the signal. Their performance was quantified by measuring signal retention and noise removal accuracy, with results reported using ROC curves across the parameter space. Additionally, we introduce a new high-resolution satellite dataset with ground truth from a real-world platform under various noise conditions, which we have made publicly available. Code, dataset, and trained weights are available at \url{https://github.com/samiarja/dvs_sparse_filter}.

Figures (8)

• Figure 1: The neuromorphic satellites observations application. (a). Satellite data is captured by a specialised mobile observatory (Astrosite) using an EVK4-HD event camera, showing an accumulated image over a 15-second recording with stars, background noise, and faint, glinting satellite signals. (b)-(c). The main goal of this paper is to benchmark all publicly available logic-based and learning-based algorithms for the noise filtering/removal task on sparse satellite data. The best-performing algorithm effectively denoises the sparse, noisy event stream while fully preserving the satellite signals. In addition, this paper proposes a new noise filtering algorithm, Cross-Convolution, and demonstrate the noise filtering capability of FEAST algorithm afshar_event-based_2020 with and without a classifier.
• Figure 2: Overview about the satellite dataset from an event camera. (a) Event stream sizes for each class, where satellites generate the fewest events, while hot pixels and background noise dominate. Faint stars blend with background noise, making them hard to distinguish. (b) Motion landscape before noise filtering, showing high contrast at zero speed, resulting in a poor Image of Warped Events (IWE). (c) Motion landscape after noise filtering, revealing distinct camera and satellite motions, resulting in a sharp IWE. (d) Example of the satellite dataset highlighting their sparseness and the surrounding hot pixels and background noise.
• Figure 3: Astrosite setup. The data was collected with Astrosite 2 deployed in Regional South Australia (Adelaide). Image adapted from marcireau2023binocular.
• Figure 4: Labelling process for the "Ev-Satellites" dataset, detailing the steps required to categorize each event as either a star, satellite, hot pixel, or noise. The goal is to augment the events by returning the per-event label array "l".
• Figure 5: Overview of the noise filtering algorithms used in this paper. (A) The pipeline of our proposed CrossConv noise filtering algorithm is categorized as a logic-based approach. It processes the event point cloud in the image space to identify high-firing rate pixels and then returns to the event space to remove them. (B)-(C) The FEAST algorithm afshar_event-based_2020, which classifies events directly during inference based on the winner neuron for each class after being trained in a supervised manner (B), and the same architecture using a linear classifier (C).