FPGA-accelerated neuromorphic vision for orbital object detection

Researchers demonstrate a real-time neuromorphic vision pipeline on FPGA hardware for tracking orbital objects — an architecture that transfers directly to the embedded seeker problem.

A preprint from mid-2026 describes an FPGA implementation of a neuromorphic event-camera processing pipeline for detecting and tracking orbital objects — satellites, debris, and reentry bodies — at rates no conventional imaging system can match. The work, FPGA-Accelerated Neuromorphic Vision System for Real-Time Orbital Object Detection (arxiv 2606.20727), is primarily aimed at space domain awareness but contains hardware architecture decisions directly relevant to airborne seeker applications.

The embedded constraint is the same

Space-based sensor payloads share the same problem as terminal seekers: severe mass, power, and volume constraints, combined with a hard latency requirement. A space sensor cannot phone home for GPU inference. A seeker payload cannot carry a server-class processor. In both cases the processing chain from sensor output to track estimate has to close on a small, qualified FPGA or SoC.

The paper demonstrates that this is achievable. Their pipeline — event representation, sparse convolutional feature extraction, and track state estimator — runs on an FPGA at 10,000+ events per second with end-to-end latency under 1 millisecond. The power envelope is under 5 watts for the processing alone, within the range of a missile or interceptor payload.

Sparse processing is the key

The architectural insight the paper relies on is that event streams are natively sparse: only pixels that see a brightness change fire, and for a slow-moving point object against a dark sky, that is a small fraction of the sensor array at any instant. Dense frame processing wastes power on pixels that contain no information. An FPGA implementation that processes only the events that arrive is a natural fit for the data structure.

The sparse convolution approach they implement routes incoming events to a fixed-function hardware block that applies a learned filter only at active pixel locations, then accumulates the result into a feature map. This avoids the memory-bandwidth bottleneck that plagues GPU inference on sparse data.

Track estimation on hardware

Beyond detection, the paper includes an on-chip Kalman filter that maintains a state estimate for detected objects between event bursts. The filter runs at the event rate rather than the frame rate, meaning the track estimate is updated with every new measurement rather than on a fixed cadence. For a fast-closing target, this makes a measurable difference in track accuracy at intercept.

Where this differs from the seeker case

The orbital detection problem has one important difference from the terminal seeker: target signatures are simpler. An orbital object against a dark sky produces a clean point source with a predictable trajectory. A UAS in the terminal phase produces a distributed, rotating signature against a cluttered, dynamic background. The FPGA architecture translates; the learned feature extractors will need to be retrained.

But the hardware template — sparse event processing, on-chip track filter, sub-watt inference at millisecond latency — is exactly what a seeker payload needs.

FPGA-Accelerated Neuromorphic Vision System for Real-Time Orbital Object Detection. arxiv 2606.20727