@article{M8C2F2DA7,
title = "Lightweight LiDAR—Camera Online Extrinsic Calibration with Multi-Dilation Encoder Blocks",
journal = "The Journal of Korean Institute of Communications and Information Sciences",
year = "2025",
issn = "1226-4717",
doi = "10.7840/kics.2025.50.5.693",
author = "Sang-Chul Kim, Yeong-min Jang",
keywords = "Camera, LiDAR, Extrinsic Calibration, Lightweight Deep Learning, KITTI odometry dataset, sensor fusion.",
abstract = "As the integration of multi-sensor systems, such as cameras and LiDAR, becomes increasingly common in various fields, the development of efficient and accurate extrinsic calibration techniques is emerging as a critical task. This article presents a novel lightweight deep-learning network for LiDAR―Camera targetless extrinsic calibration, which consists of only 4 million parameters. The proposed method utilized CNN-based multi-dilation encoder blocks which can extract multi-scale features, especially for sparse LiDAR depth image.
The proposed block allows the network to be lightweight and excel in calibration performance. The proposed method achieved translation errors of 1.08 cm, 0.18 cm, and 0.56 cm along the X, Y, and Z axes, respectively. Additionally, it achieved rotation errors of 0.182°, 0.139°, and 0.141° for roll, pitch, and yaw, respectively. The proposed method also performs calibration in a one-shot approach, which is suitable for real-time applications. These results highlight the capabilities of the proposed method in enabling reliable fusion of LiDAR and camera data, enhancing the perception capabilities of autonomous vehicles."
}