International Conferences
| 4DR P2T: 4D Radar Tensor Synthesis with Point Clouds |
| Year |
2025 |
| Month |
|
| Journal |
ITS ASIA PACIFIC FORUM 2025 |
| Author |
Woo-Jin Jung, Dong-Hee Paek, Seung-Hyun Kong* |
|
In four-dimensional (4D) Radar-based point cloud generation, clutter removal typically relies on the constant false alarm rate (CFAR) algorithm, which often fails to capture complex spatial characteristics of objects. To overcome this, we propose the 4D Radar Point-to-Tensor (4DR P2T) model, which transforms sparse radar point clouds into tensor formats that preserve environmental information without loss. Our model employs a conditional generative adversarial network (cGAN) based architecture specifically adapted for radar data. Experimental evaluations on the K-Radar dataset demonstrate the effectiveness of 4DR P2T, achieving an average PSNR of 30.39 dB and SSIM of 0.96. Leveraging this model, we conducted experiments to identify point cloud generation methods that minimize environmental information loss. Results revealed that the 5% percentile method yields the best performance, while the 1% percentile method offers the optimal balance between reduced data volume and preserved information. These results highlight 4DR P2T’s potential as a foundational approach for 4D Radar tensor generation and optimal point cloud strategies for deep learning applications. |
https://arxiv.org/abs/2502.05550
9회 연결
