A neuromorphic stereo radar framework using spinning neuron models

Small drones are hard to see: they reflect very little energy, fly low, and share the radar picture with birds and ground clutter. A single radar also measures only distance and radial speed, so it cannot place a target precisely in 3D.

The paper proposes two mobile radar nodes about 87 m apart. Each produces a range-Doppler map. Instead of passing those maps to a conventional classifier alone, the system adds a layer built from the Spinning Neuron Model (SPNM), a unit that responds to both the strength and the rhythm of its input. Rotating propellers produce a steady periodic micro-Doppler pattern; flapping wings and clutter do not. The two views are then fused to confirm the detection and triangulate its position.

The paper evaluates the approach on public 77 GHz radar recordings of drones, birds and clutter, and reports 96.8% classification accuracy with an 8.7 ms per-frame latency on an FPGA design.

A drone’s propellers spin at a nearly constant rate, which stamps a regular rhythm onto the radar echo. A bird’s wing beat and the sway of clutter are less regular, even when their average energy looks similar.

Conventional spiking networks reduce a signal to threshold-crossing events, which discards fine phase detail. SPNM keeps the signal continuous: each unit has an amplitude, an intrinsic frequency and a phase, and it adapts them so that it locks onto periodic components of its input. In the network used in the paper, the SPNM layer computes

where x is the feature vector from the range-Doppler map, W the learned weights, α sets how strongly amplitude is emphasised and β sets the resonance frequency. The sine factor makes the response oscillatory, while the amplitude factor keeps strong returns from being flattened.

With two nodes, the same target appears at slightly different positions in the two range-Doppler maps. That shift (the disparity) gives depth: z = f·B / d. Requiring both views to agree also removes echoes that appear in only one of them.

Depth accuracy gets worse quickly with distance. The paper gives the approximate error Δz ≈ z²·Δθ / B for angular resolution Δθ and baseline B, and uses it to size the baseline. Try it:

From the paper: at 500 m with Δθ = 0.1° and a 5 m error target, B ≈ 87 m. Reaching 1 km under the same conditions needs about 350 m of baseline, or coherent interferometric phase measurement and super-resolution direction-of-arrival estimation, which could cut that to 20–40 m.

Occlusion is the other reason to add nodes. A drone hidden behind terrain from one radar can still be seen by another, so the paper sketches a four-node cross or perimeter layout. The experiments in the paper use the two-node configuration.

The classifier, called StereoSPNMNet, uses two convolutional stages per radar channel, then the SPNM layer, then two fully connected layers that output micro-UAV, bird or clutter. It was trained in PyTorch on 8,000 frames (70% train, 15% validation, 15% test) of 32×32 samples per channel for 30 epochs. Input came from a public 77 GHz FMCW radar dataset of drones, birds, humans and corner reflectors (Karlsson, 2021). The left and right views are formed with a baseline-consistent range shift, a phase offset and independent noise.

Ablation — what each part contributes (paper Table 1)

Adding either stereo sensing or the SPNM layer helps on its own; the combination helps most. The paper states that all compared architectures used identical datasets, preprocessing and splits.

Against other pipelines (paper Table 2)

FPGA figures (Xilinx Zynq UltraScale+ design)

The framework is designed to grow: wider baselines or interferometric processing for longer range, more nodes for occlusion, and on-chip learning.

• Field experiments with a deployed radar pair, and hardware-in-the-loop validation on the FPGA.
• On-chip learning, so the resonance parameters adapt on the node itself.
• Autonomous baseline adjustment, where the robots reposition cooperatively to maximise localisation accuracy.
• Multi-node layouts (cross or perimeter) so a drone hidden from one radar stays visible to another.

Cite as: A. Baba, “A neuromorphic stereo radar framework using spinning neuron models,” Physical Communication 78 (2026) 103279, doi:10.1016/j.phycom.2026.103279.