Online Self-Calibration of 3D Measurement Sensors Using a Voxel-Based Network

Abstract

Multi-sensor fusion is important in the field of autonomous driving. A basic prerequisite for multi-sensor fusion is calibration between sensors. Such calibrations must be accurate and need to be performed online. Traditional calibration methods have strict rules. In contrast, the latest online calibration methods based on convolutional neural networks (CNNs) have gone beyond the limits of the conventional methods. We propose a novel algorithm for online self-calibration between sensors using voxels and three-dimensional (3D) convolution kernels. The proposed approach has the following features: (1) it is intended for calibration between sensors that measure 3D space; (2) the proposed network is capable of end-to-end learning; (3) the input 3D point cloud is converted to voxel information; (4) it uses five networks that process voxel information, and it improves calibration accuracy through iterative refinement of the output of the five networks and temporal filtering. We use the KITTI and Oxford datasets to evaluate the calibration performance of the proposed method. The proposed method achieves a rotation error of less than 0.1° and a translation error of less than 1 cm on both the KITTI and Oxford datasets.

HP3D-V2V: High-Precision 3D Object Detection Vehicle-to-Vehicle Cooperative Perception Algorithm

“…In addition, existing 3D target detection algorithms mainly use mesh-based point cloud feature extraction methods, which can be broadly categorized into 3D voxel-based and 2D column-based methods. These methodologies adopt the traditional "encoder-neck" detection architecture [20][21][22][23][24][25][26][27][28]. Voxel-based methods [20,21,25,27,28] commonly involve segmenting the input point cloud into a regular 3D voxel mesh and establishing a geometric representation across various levels through an encoder utilizing sparse 3D convolutions.…”

“…These methodologies adopt the traditional "encoder-neck" detection architecture [20][21][22][23][24][25][26][27][28]. Voxel-based methods [20,21,25,27,28] commonly involve segmenting the input point cloud into a regular 3D voxel mesh and establishing a geometric representation across various levels through an encoder utilizing sparse 3D convolutions. Following the encoder, the integration of multiscale features occurs through the neck module of a conventional 2D convolutional neural network (CNN) prior to the input entering the detection head.…”

Online LiDAR-camera extrinsic parameters self-checking and recalibration

“…It is very highly efficient to introduce semantic information to calibration sensors. Song and Lee [25] propose a method that transforms the point cloud to voxel information and uses five networks to process the voxel information and iteratively refine the final results. Ye et al [26] propose a method utilizing the geometric constraints and embedding a declarative layer into the end-to-end network.…”

Computer Vision and Machine Learning for Intelligent Sensing Systems

“…Sensors 2023, 23, 4214. https://doi.org/10.3390/s23094214 https://www.mdpi.com/journal/sensors • Song and Lee [11] studied autonomous driving and proposed a novel algorithm for online self-calibration between sensors using voxels and three-dimensional convolution kernels.…”