POINT CLOUD RECOGNITION WITH DEEP LEARNING

Abstract

This thesis studies deep learning based approaches to perform recognition tasks with 2D/3D point clouds.

Perception is the foundation of many unmanned systems such as UAVs, self-driving cars. It solves problems of "where am I", "what's around me" and "what will happen next". Sensors like LiDAR, camera are widely used to accomplish these missions. Among the many representations of measured information, point cloud is mathematically simple and effective to represent 2D/3D structures. Although point cloud is intensively applied in localization and mapping, its potential to be used in other recognition tasks is under estimated. Besides the lack of texture information, the difficulty of utilizing deep learning technique is another major reason of this phenomenon.

To fill the gap, the thesis starts with a method to solve scan matching and loop closure detection for 2D LiDAR scans, by learning scan alignment in both supervised and unsupervised fashions. Later, a general framework is proposed to process point clouds with deep network. The framework is generalizable to various applications including object classification, shape retrieval, semantic segmentation, etc. Our approaches are intensively evaluated with simulations, public datasets and real world applications, and demonstrate state-of-the-art performance.