Direction of Arrival (DOA) estimation plays a crucial role in radar perception tasks such as target detection, tracking, and imaging. The automotive industry particularly requires long-distance target detection at high speeds, necessitating real-time performance and super-resolution ability in DOA estimation. While deep learning (DL)-based methods offer unique advantages in this regard, most existing approaches suffer from grid mismatch problems, limiting their accuracy and super-resolution ability. To address this issue, a groundbreaking DL framework has been proposed in a recent publication in Science China Information Sciences.
A Novel DL Framework
The newly proposed DL framework consists of two parts, each contributing to the achievement of high-precision super-resolution DOA estimation results. The first part focuses on multi-label classification, providing initial DOA estimates on a rough grid. The second part deals with regression tasks, estimating the offset between the real DOAs and the grid based on the previous results. Additionally, skip connections are incorporated to preserve high-resolution features of the original data, enhancing the ability to discern adjacent sources. The combined outputs from both parts yield accurate and super-resolved DOA estimations.
A key contribution of the study is the consideration of the grid mismatch problem in DL-based DOA estimation. By modeling the estimation of offset values as a regression task, the proposed framework ensures that estimation is carried out on a continuous domain rather than an on-grid approach. This approach significantly enhances estimation accuracy, as it eliminates the limitations imposed by grid mismatch problems.
The novel DL framework possesses a strong resolution ability for adjacent sources. By using skip connections, high-resolution features of the original data are preserved and incorporated into the estimation process. This enhancement enables the network to effectively distinguish and accurately estimate the DOAs of closely located sources. The improved super-resolution ability of the proposed method surpasses that of traditional DL-based methods.
Evaluations and Results
The effectiveness of the DOA estimation scheme proposed in this article has been extensively evaluated using both simulation and real data. Specifically, a uniform linear array configuration with 12 elements, where the interelement distance equals half the wavelength, was employed for the experimental setup.
The first part of the experiment involved estimating the DOAs of angle pairs with fixed intervals. The proposed method demonstrated stable DOA estimation with minimal errors, showcasing its robustness and accuracy.
The second part of the experiment focused on comparing statistical characteristics among various methods. Through a series of Monte Carlo simulations, the proposed method exhibited superior performance in terms of root mean square error (RMSE), particularly under low signal-to-noise ratios (SNRs), limited snapshots, and small angle intervals. These outcomes signify the excellent adaptability and resolution ability of the proposed DL framework.
In the final part of the experiment, real data was utilized, with a substantial number of samples collected for training purposes. The trained model was then tested using a sample that was not included in the training set. The results showcased the proposed method’s capacity to accurately estimate the DOAs of closely located real targets, outperforming other methods in terms of accuracy and robustness.
The innovation presented in this article—a DL framework for DOA estimation—successfully addresses the grid mismatch problem and achieves high-precision super-resolution results. By combining multi-label classification and regression tasks, along with the preservation of high-resolution features through skip connections, the proposed method outperforms traditional approaches and surpasses other DL-based methods. The experimental evaluations, both on simulation and real data, highlight the robustness, accuracy, and adaptability of the proposed framework. The forward-thinking nature of this research promises significant advancements in DOA estimation for radar perception tasks and opens up new opportunities for applications in various industries.