Luc Tinch, University of Dayton; Nitesh Menon, Kitware, Inc.; Keigo Hirakawa, University of Dayton; Scott McCloskey, Kitware, Inc.
Keywords: bio-inspired sensors, neuromorphic sensing, computational imaging, detection, tracking, event cameras, space situational awareness
Abstract:
Event-based sensors, which capture dynamic visual environments as asynchronous event streams, have recently attracted attention in the domain of space situational awareness (SSA). Relative to traditional framing sensors, event-based sensors have SSA-relevant advantages including increased dynamic range, lower latency, and lower power consumption due to the sparse nature of the orbital environment. Increased dynamic range has been shown to enable monitoring of resident space objects (RSOs) during daytime and capture discriminating brightness variations against a dark background. In this paper, we emphasize the lower latency and higher temporal sampling of event streams, with the intention of detecting, tracking, and differentiating classes of unresolved, fast-moving, and short-lived objects. Specifically, we explore both the opportunities and challenges of event-based sensing in the context of meteor detection in a wide field of view setup, and present a dataset of real event streams captured during multiple recent meteor showers.
The dataset consists of two meteor showers – Perseids and Geminid, collected in September and December 2021 respectively. The Perseids dataset contains 31 meteors over a 2 day period in which two Prophesee VGA cameras are exclusively used with 8mm lenses. The Geminid dataset contains 29 meteors over a single day period using a Prophesee Megapixel camera with 5mm and fisheye lenses. Camera parameters (including independently-controlled positive and negative event sensitivities) were adjusted throughout the collection process to understand their impact on fast-moving, short-lived objects. To compose ground truth data, the start and end points of each meteor are annotated in 3D space. The first dataset also contains truth data for airplanes and satellites to provide information on potential distractors.
Using real-world data requires us to consider the implications of the hardware within the event-based sensors. Specifically, the presence of a fast-moving, short-lived object in a 3D space presents a situation in which the positive events and corresponding negative events are firing in such quick succession that the positive events override the negative events in their corresponding pixel location. This leads to an edge competition that must be considered in the real world dataset. As mentioned above, to compensate for the edge competition the biases of the event-based sensors are tuned to allow for higher sensitivity to positive events. While several methods of creating simulated data are possible, they do not recreate the real-world hardware issues present within the event-based sensors. Overall, real-world meteors present the most accurate data and poses a clear advantage over simulated data.
Due to the sensitivity of event-based sensors, multiple methods of filtering are required to obtain feasible meteor data. The method used separates the raw data into inceptive events, trailing events and noise. Inceptive events are defined as the first event triggered by a contrast change, and are representative of when a fast-moving, short-lived object enters a specific pixel. Trailing events are the events immediately following the inceptive event and are reflective of the intensity of the object in the 3D space. Noise is then determined by finding the inceptive events with magnitude one, i.e. contain no corresponding trailing events. Since a contrast change representing a true object in the event space will generate at least two or more trailing events, these events are considered noise. The filtering method also assigns a magnitude value to each inceptive event, representing the number of trailing events associated with the corresponding inceptive event. This allows for an intensity to also be associated with an inceptive event since the brighter the meteor, the more contrast thresholds the particular inceptive event triggers and additional trailing events are generated. Therefore the brightness of a meteor can be examined by the magnitude of the inceptive events that compose the meteor in the event space.
Date of Conference: September 27-20, 2022
Track: Optical Systems & Instrumentation