A Novel Keypoint Detection Method for Radar Point Cloud Registration

Radar point cloud registration aims to align point clouds acquired from frequency-modulated continuous-wave scanning radar at different locations to a common coordinate system, where keypoint detection is the foundation of radar point cloud registration. The keypoint detection accuracy directly impacts the precision of radar point cloud registration, consequently impacting the performance of these applications based on radar point cloud registration, such as autonomous driving, robotics, and localization. However, existing keypoint detection methods inadequately consider the effects of noise, multipath reflections, and beam spread, leading to numerous inaccurate keypoints being detected. To deal with this problem, we propose a novel keypoint detection method. First, we impose a limitation on the maximum range for keypoint detection. Then, the strongest reflection in each azimuth is regarded as candidate keypoints. Finally, we assume the candidate keypoints follow a normal distribution and utilize a threshold to select the final keypoints from the candidate keypoints. The experimental results demonstrate that the proposed method detects fewer inaccurate keypoints and achieves higher accuracy on radar point cloud registration compared to the state-of-the-art methods.