L2RDaS: Synthesizing 4D Radar Tensors for Model Generalization via Dataset Expansion

Abstract: 4-dimensional (4D) radar is increasingly adopted in autonomous driving for perception tasks, owing to its robustness under adverse weather conditions. To better utilize the spatial information inherent in 4D radar data, recent deep learning methods have transitioned from using sparse point cloud to 4D radar tensors. However, the scarcity of publicly available 4D radar tensor datasets limits model generalization across diverse driving scenarios. Previous methods addressed this by synthesizing radar data, but the outputs did not fully exploit the spatial information characteristic of 4D radar. To overcome these limitations, we propose LiDAR-to-4D radar data synthesis (L2RDaS), a framework that synthesizes spatially informative 4D radar tensors from LiDAR data available in existing autonomous driving datasets. L2RDaS integrates a modified U-Net architecture to effectively capture spatial information and an object information supplement (OBIS) module to enhance reflection fidelity. This framework enables the synthesis of radar tensors across diverse driving scenarios without additional sensor deployment or data collection. L2RDaS improves model generalization by expanding real datasets with synthetic radar tensors, achieving an average increase of 4.25\% in ${{AP}{BEV}}$ and 2.87\% in ${{AP}{3D}}$ across three detection models. Additionally, L2RDaS supports ground-truth augmentation (GT-Aug) by embedding annotated objects into LiDAR data and synthesizing them into radar tensors, resulting in further average increases of 3.75\% in ${{AP}{BEV}}$ and 4.03\% in ${{AP}{3D}}$. The implementation will be available at https://github.com/kaist-avelab/K-Radar.

4D Radar Ground Truth Augmentation with LiDAR-to-4D Radar Data Synthesis

The study presented in the paper "4D Radar Ground Truth Augmentation with LiDAR-to-4D Radar Data Synthesis" proposes a pioneering approach for ground truth augmentation tailored specifically for 4D Radar data. This work recognizes the challenges and limitations inherent in directly applying conventional ground truth augmentation (GT-Aug) methods, typically used for LiDAR, to 4D Radar data. It introduces a novel 4D Radar Ground Truth Augmentation (4DR GT-Aug) method, which synthesizes realistic radar data by leveraging a LiDAR-to-4D Radar data synthesis (L2RDaS) module. The proposed framework addresses the shortcomings of existing augmentation methodologies, particularly in maintaining the fidelity of radar-specific characteristics such as low angular resolution and the presence of sidelobes.

The L2RDaS module, central to the proposed augmentation technique, is responsible for translating augmented LiDAR point cloud data into 4D Radar data. It employs a generative adversarial network-based architecture to ensure that the synthesized radar data closely resembles actual 4D Radar measurements, incorporating measurements beyond ground truth bounding boxes (GT bboxes) and reflecting radar-specific features like sidelobes. By enhancing object density in LiDAR data and then converting it to 4D Radar data, the L2RDaS module improves object detection accuracy, as demonstrated in experiments conducted on the K-Radar dataset.

Experimental Results and Performance

The study rigorously evaluates the effectiveness of the proposed 4DR GT-Aug method using the K-Radar dataset, which offers unique access to 4D Radar tensor data. It utilizes RTNH, a baseline model designed for 4D Radar-only object detection, to benchmark the performance of the proposed augmentation strategy. The results illustrate a notable increase in object detection performance: the method achieves a 4.8 percentage point improvement in Bird’s Eye View (BEV) and a 1.54 percentage point boost in 3D detection accuracy over the no-augmentation baseline. Moreover, the 4DR GT-Aug method also surpasses the conventional GT-Augantation practices applied to 4D Radar by 1.54 percentage points in BEV and 0.61 percentage points in 3D object detection.

Contributions and Implications

This paper makes significant contributions to the field of autonomous driving and sensor data augmentation:

• L2RDaS Development: The introduction of the L2RDaS for 4D Radar data synthesis is a noteworthy advancement. It provides a method for generating radar tensor data from LiDAR point clouds, capturing both GT bbox data and the surrounding radar characteristics.

• 4DR GT-Aug Methodology: By integrating the L2RDaS module, the study establishes the first GT-Aug method specifically attuned to 4D Radar, which incorporates radar characteristics like sidelobes and low angular resolution in the augmentation process.

• Empirical Validation: Through comprehensive experiments, the effectiveness of the 4DR GT-Aug approach is validated, showing improved detection performance and demonstrating the practical utility of synthesized radar data in training object detection models.

The enhancement offered by 4DR GT-Aug in synthesizing realistic radar training data is anticipated to significantly impact radar-based perception systems in autonomous vehicles. Its ability to improve object detection accuracy even in challenging environments highlights its potential applications in enhancing the robustness of autonomous navigation systems. Future research may explore the integration of temporal dynamics and further refinement of synthetic radar data quality, addressing performance challenges in adverse conditions for broader applicability.

In conclusion, this study provides a robust framework for 4D Radar data augmentation, underscoring the importance of accommodating radar-specific characteristics in data synthesis for improving object detection in autonomous driving.