I proposed a robust and complementary visual-4D radar odometry for dynamic environments, such as urban areas. Monocular visual odometry suffers from the metric scale ambiguity issue and vulnerability to moving features. The Dynamic Object Removal Masks (DORM) were suggested for rejecting moving features by drawing circles around the points projected to the query image from 4D radar Doppler velocity measurements. As shown in video, it can be observed that the proposed method exhibits better robustness in dynamic environments compared to conventional 5-point RANSAC visual odometry. Additionally, the scale ambiguity issue was addressed through ego velocity estimation from 4D radar. The performance of the proposed method has been validated to outperform other SLAM techniques through real-world implementation experiments. Please refer to our paper for detailed methodology and results.  [paper link] [youtube]

Most vector mapping studies utilize IPM (Inverse Perspective Mapping), which requires knowledge or estimation of the ground plane information. While ground plane estimation can be achieved from 4D radar data, it is not accurate enough for IPM due to noise caused by multi-path reflections. Therefore, we conducted ground plane correction based on vanishing point tracking. Our proposed method showed quantitatively better performance compared to conventional methods using LiDAR-based ground plane estimation as a pseudo-ground truth. As depicted in the figure, the generated vector maps are sufficient for leveraging loop matching, which will be detailed in the next section. We performed real-time mapping of a 3D vector map using IPM from lane masks detected from 2D images. Please refer to our paper for detailed methodology and results. [paper link] [youtube]

Loop detection for pose graph optimization is crucial part of SLAM for minimizing trajectory drift. However, relying soley on vector maps is insufficient for loop detection due to their similar duplicated shapes. Therefore, additional 4D radar data will be utilized for loop detection, given its wider FOV. Because 4D radar only looks forward unlike LiDAR, we adopted Z projection image for encoding 4D radar data into data base. We introduced double-check loop detection using both 4D radar (coarse) and vector map (fine), followed by matching with the vector map for elaborate pose graph optimization. Experimental results validated that the proposed double check loop detection shows better performance than conventional strategies. Please refer to our paper for detailed methodology and results.  [paper link] [youtube]