Myles Fernau, Southwest Research Institute; Camden Ertley, Southwest Research Institute; David Anthony, Southwest Research Institute; Chau Ton, Southwest Research Institute; Christopher Comeau, Southwest Research Institute
https://doi.org/10.64861/HDIM5190
Keywords: Neuromorphic computing, Spiking neural networks, Event cameras, Space Domain Awareness
Abstract:
Detecting and tracking small, low-signature objects from space and high-altitude platforms remains a central challenge in Space Situational Awareness (SSA) and Space Domain Awareness (SDA). Future architectures emphasize distributed constellations, onboard autonomy, and reduced size, weight, and power (SWaP), yet conventional electro-optical (EO) detection pipelines rely on GPU-accelerated convolutional neural networks that often consume tens of watts and struggle to detect objects occupying only a few pixels. This paper presents the early-stage development of a neuromorphic sensing and processing architecture designed to reduce SWaP while preserving detection fidelity for extremely small targets._x000D_
_x000D_
The goal of this research is to investigate whether spiking neural networks (SNNs) running on neuromorphic processors, coupled with both frame-based and event cameras, can provide a viable low-power alternative to traditional GPU-based pipelines for small-object detection in SSA-relevant scenarios. The target regime includes objects as small as a few square pixels, moving at high apparent angular rates, where appearance-based detection methods are ineffective._x000D_
_x000D_
The methodology follows a staged development approach. First, a convolutional neural network baseline is implemented using a U-Net style encoder-decoder architecture trained on publicly available datasets relevant to small uncrewed aerial systems and resident space objects. This baseline establishes detection precision, recall, and latency on embedded GPU hardware. The model is then quantized using quantization-aware training to approximate the reduced precision constraints typical of neuromorphic hardware and to establish an apples-to-apples comparison point._x000D_
_x000D_
Second, the quantized convolutional model is converted to an SNN representation targeting the Brainchip Akida neuromorphic architecture. The conversion leverages automated CNN-to-SNN mapping tools, followed by iterative refinement to address performance degradation associated with non-differentiable spiking activations and low-bit integer weights. Performance is evaluated in software emulation and FPGA-based implementations, with metrics including precision, recall, execution speed, and estimated resource utilization as a surrogate for power._x000D_
_x000D_
Finally, the sensing modality is transitioned from frame-based cameras to event cameras, resulting in a fully neuromorphic stack. Event cameras provide asynchronous one-bit intensity changes with high dynamic range and no motion blur, which is attractive for fast-moving space objects but requires spatio-temporal processing. The encoder portion of the network is modified to incorporate depthwise-separable spatio-temporal convolutional blocks compatible with neuromorphic execution. Temporal window parameters are swept over controlled target-speed regimes to characterize the trade space between latency and detection fidelity._x000D_
_x000D_
Because the program will be in its first months at the time of manuscript preparation, this paper will emphasize architectural design decisions, early baseline performance results, and preliminary CNN-to-SNN conversion outcomes. Initial comparisons are expected to quantify the detection accuracy gap between conventional and neuromorphic implementations and to estimate achievable reductions in computational resources. Anticipated performance targets include maintaining greater than 80 percent precision and recall for neuromorphic implementations while reducing power consumption relative to embedded GPU baselines._x000D_
_x000D_
The applicability to SSA and SDA lies in enabling onboard detection and tracking of small, dim objects under strict SWaP and thermal constraints. A successful neuromorphic pipeline would support distributed space-based sensors, hosted payloads, high-altitude platforms, and responsive space missions where conventional GPU solutions are infeasible. By systematically characterizing the performance, resource, and latency trade space across conventional, quantized, and fully neuromorphic architectures, this work aims to provide the SSA community with a clear assessment of whether neuromorphic sensing represents a practical path toward next-generation autonomous space surveillance systems._x000D_
Date of Conference: September 15-18, 2026
Track: SDA Systems & Instrumentation