Vithurshan Suthakar, York University; Regina Lee, York University; Youssef Korny, York University
https://doi.org/10.64861/SIZJ7465
Keywords: Event-based imaging,Neuromorphic vision,Space situational awareness (SSA), Resident space objects (RSOs), Onboard processing, Small Aperture, Wide Field of View Optical Systems
Abstract:
Event-based (neuromorphic) cameras encode visual information as an asynchronous stream of per-pixel brightness-change events rather than fixed-rate intensity frames, giving microsecond temporal resolution, high dynamic range, and low bandwidth and power. These properties suit the sparse, time-driven regime of space situational awareness (SSA), and particularly the observation of resident space objects (RSOs) transiting at high apparent angular rates. Event cameras have been shown to support ground-based SSA in daytime and nighttime without changing optics, as well as star-based attitude estimation, which motivates dual-purpose payloads for both navigation and SSA.
This study presents E-STAR (Event-based Star and RSO Imager), an event-based payload for star and RSO observation in a stratospheric environment, built to establish the feasibility of continuous, high-frequency imaging at low data rates under flight conditions relevant to future orbital deployment. It pairs a Prophesee EVK4 event camera (Sony IMX636, 1280 by 720, 4.86 micrometer pitch, approximately 120 decibel dynamic range) with a ZEISS Dimension 2 by 25 lens (25 millimeter, f-number 2, less than 0.1 percent distortion) previously flown and qualified on stratospheric balloons. This gives a 14.2 by 8.0 degree field of view at approximately 40 arcseconds per pixel, with onboard processing on a Raspberry Pi. E-STAR is planned to fly on a CNES/CSA zero-pressure balloon campaign from Palmas, Brazil, in 2027, a gondola platform that supplies regulated power, telemetry, and payload recovery.
The concept of operations runs two concurrent onboard paths. A storage path captures the event stream continuously to timestamped HDF5 and losslessly compresses it, preserving a complete raw record for post-flight validation. In parallel, a detection path applies an event-native, density-driven clustering pipeline that separates star-like and RSO-like activity from background and produces candidate tracks and compact metadata at a fraction of the raw volume. This decoupling gives a low-rate science product while retaining every event needed to verify it.
Developing this pipeline on the Event-Based Space Situational Awareness (EBSSA) dataset, we identify a limitation in how such methods are assessed: compression ratio does not measure whether object signal is preserved, since the two are independent. We introduce a signal-preservation metric based on labelled object trajectories, with a two-condition acceptance test that requires both a signal-retention floor and better-than-random selectivity. Under this metric, the standard fixed-threshold neighbour denoiser behaves as a global rate-burst gate rather than a spatial-structure detector, and the event-rate distribution is strongly over-dispersed, which invalidates the Poisson-based significance thresholds in common use. We therefore develop a rate-normalised density filter that scales the retention threshold by the instantaneous event rate; on representative data it substantially improves signal retention and purity, and an ablation shows that the rate normalisation, not the neighbourhood size, is responsible. We also examine a learned detector, a recurrent vision transformer, as an offline comparison that quantifies the trade between deterministic flight-safe processing and higher-capacity learned detection.
The flight is designed to assess star-field detectability under jitter, the persistence and separability of RSO tracks, and the viability of low-power onboard processing. We estimate sensitivity by aperture-scaling from published event-camera results: the predicted limiting magnitude is near 6 for the 12.5-millimeter pupil under quiet pointing, and about one magnitude brighter during rotation bursts. From attitude and GPS logs of the night-time segments of two prior CSA/CNES Strato-Science flights, we find the uncontrolled gondola stays below the rate at which event cameras cease to detect stars for roughly 60 percent of the time, interrupted by faster rotation bursts; active pointing raises this to nearly full-time. We treat this rotation as the dominant operational risk and use the measured distribution to predict per-epoch detectability. The work delivers a characterized payload and a validated processing workflow for continuous star and RSO imaging at low data rates, and a defined flight campaign toward orbital neuromorphic SSA.
Date of Conference: September 15-18, 2026
Track: SDA Systems & Instrumentation