Abstract:Automotive Multiple-Input Multiple-Output (MIMO) radars have gained significant traction in Advanced Driver Assistance Systems (ADAS) and Autonomous Vehicles (AV) due to their cost-effectiveness, resilience to challenging operating conditions, and extended detection range. To fully leverage the advantages of MIMO radars, it is crucial to develop an Angle of Arrival (AOA) algorithm that delivers high performance with reasonable computational workload. This work introduces AAETR (Angle of Arrival Estimation with TRansformer) for high performance gridless AOA estimation. Comprehensive evaluations across various signal-to-noise ratios (SNRs) and multi-target scenarios demonstrate AAETR's superior performance compared to super resolution AOA algorithms such as Iterative Adaptive Approach (IAA). The proposed architecture features efficient, scalable, sparse and gridless angle-finding capability, overcoming the issues of high computational cost and straddling loss in SNR associated with grid-based IAA. AAETR requires fewer tunable hyper-parameters and is end-to-end trainable in a deep learning radar perception pipeline. When trained on large-scale simulated datasets then evaluated on real dataset, AAETR exhibits remarkable zero-shot sim-to-real transferability and emergent sidelobe suppression capability. This highlights the effectiveness of the proposed approach and its potential as a drop-in module in practical systems.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to estimate the Angle of Arrival (AoA) of radar signals efficiently and accurately in Advanced Driver - Assistance Systems (ADAS) and Autonomous Vehicles (AV). Specifically, the paper proposes a grid - free AoA estimation method based on Transformer - AAETR (Angle of Arrival Estimation with TRansformer), aiming to overcome the limitations of traditional methods (such as the Iterative Adaptive Approach (IAA)) in terms of high computational cost, strong dependence on grid resolution, and performance degradation across Signal - to - Noise Ratio (SNR). ### Main contributions of the paper: 1. **Proposing a fully differentiable Transformer - based AoA model**: This model significantly outperforms existing super - resolution methods. By adopting a sparse detection architecture, AAETR achieves grid - free estimation, thus maintaining a low computational cost even at high grid resolutions. 2. **Demonstrating zero - shot transfer ability from simulated data to real data**: AAETR performs excellently on unseen real - world driving - scene data, highlighting its robustness and adaptability in practical applications. 3. **Proposing a new evaluation method**: This method correlates the performance of downstream tasks (such as 3D object detection and segmentation) with spectral estimation accuracy, providing a more comprehensive evaluation of the model's practicality. ### Background and motivation: - **Advantages of MIMO radar**: Compared with LiDAR, MIMO radar has the advantages of high cost - effectiveness, strong anti - bad - weather ability, and wide detection range. - **Limitations of traditional AoA algorithms**: Traditional AoA algorithms (such as MUSIC, IAA, etc.) perform well under certain conditions, but have problems such as high computational complexity and the need for multiple time snapshots, which limit their application in real - time vehicle - borne radar systems. - **The rise of deep - learning methods**: In recent years, deep - learning - based AoA algorithms have begun to emerge. These methods can achieve faster inference speeds while maintaining high performance. However, existing research mainly focuses on end - to - end perception tasks or performance improvement in limited synthetic scenarios, lacking evaluation of real - world performance and robustness. ### Method overview: - **Model architecture**: AAETR draws on the idea of DETR (End - to - End Object Detection with Transformers) and contains an encoder and a decoder. The encoder processes array signals, and the decoder predicts targets through the cross - attention mechanism. - **Loss function**: It is trained using a modified set - prediction loss. This loss function finds the optimal permutation through bipartite matching to minimize the difference between the prediction and the real label. - **Evaluation framework**: AAETR is comprehensively evaluated from three aspects: detection ability, measurement accuracy, and algorithm robustness, especially in multi - target detection and under different SNR conditions. ### Experimental results: - **Performance on synthetic datasets**: AAETR shows better performance than IAA in multi - target detection and under different SNR conditions. - **Zero - shot transfer ability on real datasets**: Although AAETR is only trained on synthetic data, it still performs excellently on real - world data, especially in sidelobe suppression. ### Conclusion: As a grid - free AoA estimation method based on Transformer, AAETR not only performs well on synthetic datasets but also has zero - shot transfer ability from simulated data to real data. This makes AAETR a potential module that can be integrated into actual vehicle - borne radar systems to enhance the perception ability of ADAS and AV.

Angle of Arrival Estimation with Transformer: A Sparse and Gridless Method with Zero-Shot Capability

Efficient Gridless DoA Estimation Method of Non-uniform Linear Arrays with Applications in Automotive Radars

Deep Learning Based Computationally Efficient Unrolling IAA for Direction-of-Arrival Estimation

Self-Supervised Learning for Enhancing Angular Resolution in Automotive MIMO Radars

Antenna Failure Resilience: Deep Learning-Enabled Robust DOA Estimation with Single Snapshot Sparse Arrays

DeepAoANet: Learning Angle of Arrival From Software Defined Radios With Deep Neural Networks

Redefining Automotive Radar Imaging: A Domain-Informed 1D Deep Learning Approach for High-Resolution and Efficient Performance

A Deep Learning Based AoA Estimation Method in NLOS Environments

Angular Super-Resolution Algorithm for Millimeter-Wave Automotive Radar Based on Virtual Array Augmentation

Generalized Optimization of Sparse Antenna Arrays for High-Resolution Automotive Radar Imaging

TransRadar: Adaptive-Directional Transformer for Real-Time Multi-View Radar Semantic Segmentation

A Two-Stage Multi-Layer Perceptron for High-Resolution DOA Estimation

Efficient tensor model‐Toeplitz matrix iterative reconstruction for angle estimation with nested array

Time-Varying Angle Estimation of Multiple Unresolved Extended Targets via Monopulse Radar

A Deep Learning Approach for Doppler Unfolding in Automotive TDM MIMO Radar

4D Automotive Radar Sensing for Autonomous Vehicles: A Sparsity-Oriented Approach.

High-Precision DOA Estimation Based on Synthetic Aperture and Sparse Reconstruction

Deep-Neural-Network-Enabled Vehicle Detection Using High-Resolution Automotive Radar Imaging

4D Automotive Radar Exploiting Sparse Array Optimization and Compressive Sensing

TempEE: Temporal–Spatial Parallel Transformer for Radar Echo Extrapolation Beyond Autoregression

AODet: Aerial Object Detection Using Transformers for Foreground Regions