Scene understanding from 3D point clouds and RGB images for autonomous driving

• Type: Thesis
• Date: 2021-07-22

Authors

Rolando de Sousa Chichorro Avides Moreira (Author) FEUP

Thesis

Scene understanding from 3D point clouds and RGB images for autonomous driving

Master's Degree

Dissertation

Scientific classification

FOS: Engineering and technology > Electrical engineering, Electronic engineering, Information engineering

Other information

Resumo (PT): Autonomous cars are often equipped with 3D data acquisition sensors and devices, e.g., LiDAR, which provide a 3D point cloud that describes the surroundings. Direct acquisition of 3D data from these sensors is commonly used for obstacle avoidance and mapping. Analysing 3D point clouds is complex since point clouds are unstructured, unordered, and contain a varying number of points. The most common approach used for scene understanding in images is the Convolutional Neural Network. Although CNNs achieve high performance in image analysis, they cannot be applied naturally on point clouds. Several methods for extending CNNs to 3D point cloud analysis have been proposed, such as rasterization into a 3D voxel grid to use directly a CNN or using a Graph Convolutional Network. The main goal of this dissertation is to study and compare different approaches for scene understanding from 3D point clouds within the scope of driving automation systems. Moreover, the project contemplates the study of sensor fusion approaches, namely how to combine 3D point clouds and images. In light of this, this project uses a sensor fusion technique called pointpainting, which uses images segmentation to enhance 3D object detection on point clouds.

Language: English

No. of pages: 92