PCRMLP: A Two-Stage Network for Point Cloud Registration in Urban Scenes

Area of Science:

• Computer Vision
• Robotics
• Geomatics

Background:

• Point cloud registration is vital for 3D mapping and localization.
• Urban scenes present challenges like large data, similar environments, and dynamic objects.
• Instance-level localization offers a more intuitive approach.

Purpose of the Study:

• To propose PCRMLP, a novel Multilayer Perceptron model for urban scene point cloud registration.
• To develop an instance-level urban scene representation for robust feature extraction and transformation estimation.
• To achieve comparable or superior registration performance to existing learning-based methods.

Main Methods:

• Leveraging semantic segmentation and DBSCAN for instance-level urban scene representation and descriptor generation.
• Employing a lightweight Multilayer Perceptron network in an encoder-decoder architecture for implicit transformation estimation.
• Integrating an Iterative Closest Point (ICP) refinement module for enhanced accuracy.

Main Results:

• PCRMLP achieves satisfactory coarse transformation estimates from instance descriptors in 0.0028 seconds on the KITTI dataset.
• The method demonstrates robust feature extraction and dynamic object filtering.
• With ICP refinement, PCRMLP yields a rotation error of 2.01° and a translation error of 1.58 m, outperforming prior learning-based approaches.

Conclusions:

• PCRMLP offers a promising approach for efficient and accurate coarse registration of urban scene point clouds.
• The instance-level representation method enables robust and semantically meaningful registration.
• The findings support PCRMLP's application in instance-level semantic mapping and localization.