Automotive Radar DoA Estimation Using Structured ADMM-Net

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional automotive radar systems face challenges in direction of arrival (DoA) estimation due to high computational requirements and reliance on prior assumptions about data distributions, leading to suboptimal performance and increased complexity, which is not suitable for real-time applications.

Innovation Solution

The use of an ADMM-Net neural network with hermitian-centrohermitian or circulant matrix constraints to process radar signals, reducing computational complexity and eliminating the need for a priori information about object distribution, thereby improving robustness and efficiency in DoA estimation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional signal processing methods are used for DoA estimation, then measurement precision can be achieved, but device complexity and computing power requirements increase significantly

Engineering Contradiction:
Improvedirection of arrival estimation accuracyVSAvoidcomputer and memory requirements
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces conventional mechanical signal processing methods (FFT-based spectral analysis, iterative optimization algorithms) with a neural network system. The neural network learns optimal signal processing transformations during training and applies them during inference, substituting complex computational mechanics with learned patterns that achieve similar or better measurement precision with reduced computational overhead.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The neural network is trained offline using labeled radar data to learn optimal signal processing strategies. This preliminary training phase allows the network to internalize complex processing patterns, so that during actual DoA estimation, the pre-learned knowledge is applied directly without requiring complex real-time computations, thus reducing device complexity while maintaining precision.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If conventional DoA estimation algorithms are implemented, then direction of arrival can be determined, but runtime is increased and real-time processing becomes difficult

Engineering Contradiction:
ImproveDoA estimation accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The neural network performs computationally intensive learning during an offline training phase, where complex patterns are discovered and stored as learned weights and biases. During actual DoA estimation, only forward propagation through the trained network is required, which is computationally efficient and enables real-time processing while maintaining high accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The neural network creates a simplified computational model that copies the essential processing patterns of complex conventional algorithms. Instead of executing multiple iterative optimization steps or large FFT operations in real-time, the network uses a compact set of learned parameters that replicate the computational outcomes, dramatically reducing runtime while preserving measurement precision.

Inventive Principle:
Principle #26Copying

3Device complexity

If prior assumptions about data distribution are used in DoA estimation, then computational complexity is reduced, but measurement precision deteriorates due to suboptimal performance

Engineering Contradiction:
Improvecomputational complexityVSAvoidDoA estimation accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The neural network provides a dynamic solution that adapts to different data distributions without requiring explicit prior assumptions. During training, the network learns the actual data distribution characteristics from the training set, and during inference, it automatically adjusts its processing based on the learned patterns. This dynamic adaptation eliminates the need for fixed prior assumptions while maintaining low computational complexity and high measurement precision.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4707854A1Structured neural network for automotive radar direction of arrival estimation
Publication Date: 2026.03.11 NXP USA INC
  • EP4707854A1 patent drawingFigure 1A
  • EP4707854A1 patent drawingFigure 1B
  • EP4707854A1 patent drawingFigure 2

AI summary

In an automotive radar system, a measurement vector is determined using signals received a plurality of radar receiver modules. An expression is determined that defining an iteration of an optimization problem configured to determine an optimized output amplitude vector based on the measurement vector, wherein the expression includes a first parameter that is a hermitian-centrohermitian matrix or a circulant matrix. The automotive radar system includes a neural network and each node of the neural network solves iterations of the expression to determine an optimized value of the first parameter. A final node of the neural network determines the optimized output amplitude vector based on the optimized value of the first parameter and an estimated direction of arrival of a first object is determined using the optimized output amplitude vector.