Development of a single-object sim-to-real neural network for 3D object detection using point clouds
Year: 2023
Type: dissertation
Abstract: Perception plays a critical role in the operation of autonomous cars. Neural networks and specifically deep learning have emerged as a prominent technology for object identification in the immediate vicinity of autonomous vehicles. However, their effectiveness relies on extensive and expensive labeled datasets required for supervised training. While simulated data offers a potential alternative, neural networks trained solely on simulated data fail to achieve comparable performance to networks trained with real-world data. The objective of this thesis is to develop a neural network capable of converting synthetic data generated in simulations to real-world data. This thesis focuses on the domain adaptation of 3D point clouds obtained from LiDAR sensors, as opposed to 2D photos, that already have some actual working solutions as CycleGAN. By addressing this objective, the research aims to bridge the gap between simulated and real- world data for 3D point clouds, enabling more accurate and robust perception capabilities for autonomous vehicles. The following points are to be processed by Mr. Sergi Flo Salinas: - Scientific research and presentation of the state of the art of LiDAR 3D domain adaptation and SeedFormer network architecture - Development of a deep learning algorithm for adapting simulated point clouds to reality using SeedFormer network. - Documentation of the results The elaboration should document the individual work steps in a clear form. The candidate undertakes to carry out the master's thesis independently and to indicate the scientific aids used. The submitted work remains the property of the chair as an examination document
Author: Flo I Salinas, Sergi
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7018484825
