Point cloud based place recognition and localization for autonomous robots using deep learning

Abstract

Localization is of paramount importance for robots such as self-driving cars and drones to achieve full autonomy, especially when GPS signals are unavailable.

Most of the researchers focus on putting forward 2D-images-based place recognition and localization algorithms, few researchers pay attention to letting robots use point cloud to perform localization work, due to the difficulty of extracting point cloud's local features, which can be transformed into global descriptors. However, point cloud do outperform images in many aspects: point cloud are invariant to drastic lighting changes, thus making it more robust to perform localization tasks on queries that are taken from different times of the day or different seasons of the year. Also, more accurate localization information can be obtained from point cloud compared to images due to the availability of precise depth information in point cloud.

To learn a robust point-cloud-based representation. We propose a novel attentional encoding strategy named Pointnet-ShadowVLAD. In this work, our contribution lies in: (1) Introducing non-informative clusters to help the network ignore misleading information in the point cloud. (2) Propose a novel initialization approach to set the location of informative clusters and non-informative clusters, thus the attention of the network can be optimized from both topological priors and information-based learning.