The paper proposes an automated framework for registering prior point clouds with roadside camera images. The key components of the framework are:
Neighbor Rendering: An efficient rendering method that generates realistic grayscale views from the point cloud while preserving the 2D-3D correspondence, addressing the challenges of sparse and uneven point cloud distribution in roadside scenes.
Initial Guess Estimation: A pipeline that estimates the initial guess for the extrinsic parameters using only rough guesses of the camera's position, reducing the dependency on accurate initial configurations.
Extrinsic Parameters Optimization: A procedure that extracts line features from both the generated views and camera images using the Segment Anything Model (SAM), and optimizes the extrinsic parameters by minimizing the reprojection error of these line features.
The proposed method is evaluated on a self-collected dataset of roadside scenes, achieving an average translation error of 0.079m and a rotation error of 0.202°. The registered image-point cloud pairs are further applied to the task of roadside 3D object detection, demonstrating the practical effectiveness of the framework.
To Another Language
from source content
arxiv.org
Djupare frågor