Direction of arrival (DOA) estimation using circular convolutional network

A device comprising:

• a signal processing unit that includes processing circuitry and memory coupled to the processing circuitry, wherein the processing circuitry includes multiple vector processing units, each vector processing unit configured to:
• receive an antenna input vector that is a representation of radar return signal amplitudes received by each antenna of an array of antennas;
• receive an angular spectrum vector that indicates a likelihood of a return signal originating from ranges of angles corresponding to each component of the antenna input vector;
• retrieve a first weighting vector and a second weighting vector for each vector processing unit of the multiple vector processing units from the memory;
• generate a processed antenna input vector by performing a circular convolution of the antenna input vector with the first weighting vector;
• generate a processed angular spectrum vector by performing a circular convolution of the angular spectrum vector with the second weighting vector;
• and
• generate a refined angular spectrum vector by applying a non-linear activation function to a sum of the processed antenna input vector and the processed angular spectrum vector, wherein the non-linear activation function specifies a non-linear mapping between input vector component values of the processed antenna input vector and output vector component values of the refined angular spectrum vector;
• wherein the processing circuitry is configured to output a final angular spectrum vector that indicates an estimated angular position of one or more radar targets with respect to the array of antennas based on the refined angular spectrum vectors generated by each of the vector processing units.

A method for estimating a direction of arrival (DOA) from radar signals, the method comprising:

• at vector processing units of a signal processing unit that includes processing circuitry and memory coupled to the processing circuitry and that is connected to a radar transceiver:
• receiving an antenna input vector that is a representation of radar return signal amplitudes received by each antenna of an array of antennas;
• receiving an angular spectrum vector that indicates a likelihood of a return signal originating from ranges of angles corresponding to each component of the antenna input vector;
• retrieving a first weighting vector and a second weighting vector for one of the vector processing units of the vector processing units from the memory;
• at a circular convolution engine, generating a processed antenna input vector by performing a circular convolution of the antenna input vector with the first weighting vector;
• at the circular convolution engine, generating a processed angular spectrum vector by performing a circular convolution of the angular spectrum vector with the second weighting vector;
• and
• at an activation function engine, generating a refined angular spectrum vector by applying a non-linear activation function to a sum of the processed antenna input vector and the processed angular spectrum vector, wherein the non-linear activation function specifies a non-linear mapping between input vector component values of the processed antenna input vector and output vector component values of the refined angular spectrum vector;
• wherein the processing circuitry outputs a final angular spectrum vector that indicates an estimated angular position of one or more radar targets with respect to the array of antennas based on refined angular spectrum vectors generated by each of the vector processing units.