Implementation of a Spiking Neural Network solving the 1D and 2D Discrete Fourier Transform and the OS-CFAR detection algorithm for automotive radar raw data.
This project shows a proof-of-concept of an SNN simulation able to replace the Fourier transform and object detection for radar data. The focus is not to offer an optimized implementation in neuromorphic hardware, so performance parameters like energy consumption and latency are not measured and optimized. Be also aware that the 2D implementation consumes a lot of memory resources and some computers might not be able to run it.
The algorithm has been introduced in the following paper:
López-Randulfe, J., Duswald, T., Bing, Z., and Knoll, A. "Spiking Neural Network for Fourier Transform and Object Detection for Automotive Radar". Frontiers in Neurorobotics (2021) doi: 10.3389/fnbot.2021.688344.
Once you are in the main project directory, just install it as a normal Python3 package:
pip install .
The main module is the entry point for testing the algorithm. Create your custom configuration for the experiment or use the custom one, and select the number of dimensions (1 or 2) and the method to be used ("numpy" for the standard DFT and OS-CFAR, or "SNN" for the spiking version):
usage: main.py [-h] [-d {1 2}] [-m {numpy SNN}] [-f] config_file
positional arguments:
config_file Relative location of the configuration file
optional arguments:
-h, --help show this help message and exit
-d , --dimensions {1 | 2} number of DFT dimensions
-m , --method {numpy | SNN} method used for running the system
-f [] Get the S-DFT data from a local file
-s [] Show the plot after the simulation
Example of usage:
python3 main.py config config/scenario1_default.json --dimensions=1 --method=SNN -s
There is data containing a sample scenario with three targets and typical noise
sources in automotive radar. For more detailed specifications of the scene, go
to the json file in data/BBM/scenario1.
The parameters for the experiments are specified in a config file in json format. There is a default configuration in the config folder. Feel free to modify it and try different simulation parameters:
- cfar_args: Standard OS-CFAR algorithm parameters. The guard and neighbour
cell parameters refer to half-window size. This means that
guarding_cells=3will create a window of 6 cells around the central cell for the 1D case, and a window of 6x6 cells for the 2D case. zero_padding indicates whether to calculate the CFAR on the borders by adding zeroes around the frame - cfar_encoding_parameters: Parameters for implementing time coding on the provided input of the S-OSCFAR. NOTE: x_max is automatically re-calculated within the code, based on the range of the input data
- dft_encoding_parameters: Paramters for implementing the rate coding of the S-DFT.