The project developed a neuromorphic processing pipeline for automotive radar data that replaces conventional digital signal processing stages with spiking neural networks. The first step in the pipeline is the conversion of the analog radar return into a stream of spikes. A custom integrate‑and‑fire circuit replaces the standard ADC and immediately encodes the incoming signal into temporal pulses that can be fed directly to neuromorphic hardware. The subsequent Fourier transform, normally performed by an FFT block, is implemented as a time‑coded spiking network whose connectivity and synaptic weights are derived analytically from the Fourier equations. Two variants of the constant false alarm rate (CFAR) detector were realised on the SpiNNaker 2 chip: an oscillatory‑spike CFAR (OS‑CFAR) developed by the Technical University of Munich and a conventional CFAR (CA‑CFAR) from the Technical University of Dresden. Benchmarks show that the outputs of both spiking CFAR implementations match the results of the classical algorithms exactly, confirming the functional correctness of the neuromorphic approach. The efficiency of the hardware remains to be fully quantified, as the SpiNNaker 2 platform is still under development.

For the object‑data processing stage, two distinct networks were trained and evaluated. A clustering network based on time‑coded inputs and unsupervised learning was able to group radar points with an accuracy of roughly 60 %. The network is lightweight, generating only two spikes per input value (representing x and y coordinates). A continuous attractor network, using rate‑coded values, was designed for tracking dynamic objects. It achieves a 79 % detection rate and a positional error of 3.22 m, while running at 100 frames per second on a SpiNNaker 1 board. These results demonstrate that neuromorphic architectures can perform both detection and tracking in real time with competitive accuracy.

Data management was addressed by selecting the HDF5 format for its stability, high read/write performance, and cross‑platform support. Experiments with 5 GB datasets revealed that HDF5 offers the fastest write speed, whereas the Feather format provides the best read speed but was excluded due to lack of a stable release. The consortium also developed a lightweight software layer that allows partners to load and process data from the servers with a few commands, ensuring that all collaborators work with identical, up‑to‑date test sets.

The consortium comprises the Technical University of Munich, the Technical University of Dresden, Infineon AG, BMW AG, and the OTH university. TUM led the design of the neuromorphic signal‑to‑spike conversion, the spiking Fourier and CFAR modules, and the unsupervised learning networks. TUD supplied the SpiNNaker 2 hardware and implemented the CFAR variants. Infineon and BMW contributed expertise on the conventional radar processing chain and the high‑resolution radar hardware. Together they defined the modular system architecture, which includes a high‑resolution radar, an Ethernet switch, a SpiNNaker board with ARM processors and FPGA fabric, and a PC for higher‑level processing. The project ran over its full duration, with all partners collaborating closely on data structure, software development, and algorithm evaluation. The consortium’s findings were published in a joint paper on the state of the art and challenges of automotive radar processing with spiking neural networks.