This master thesis proposal involves developing a method for 3D localization from dense RGB-D data and dedicated maps. Due to the large amount of data from high-resolution sensors, traditional deterministic techniques like ICP are infeasible. The goal is to explore new approaches to make fast, high-density RGB-D localization more general and applicable in practice. Activities will include a literature review, implementation with simulation and real data, developing map contents and format, and generalizing the method for arbitrary maps. The position requires relevant coursework, programming skills in Keras/TensorFlow, and experience with AI algorithms like neural networks.
This master thesis proposal involves developing a method for 3D localization from dense RGB-D data and dedicated maps. Due to the large amount of data from high-resolution sensors, traditional deterministic techniques like ICP are infeasible. The goal is to explore new approaches to make fast, high-density RGB-D localization more general and applicable in practice. Activities will include a literature review, implementation with simulation and real data, developing map contents and format, and generalizing the method for arbitrary maps. The position requires relevant coursework, programming skills in Keras/TensorFlow, and experience with AI algorithms like neural networks.
This master thesis proposal involves developing a method for 3D localization from dense RGB-D data and dedicated maps. Due to the large amount of data from high-resolution sensors, traditional deterministic techniques like ICP are infeasible. The goal is to explore new approaches to make fast, high-density RGB-D localization more general and applicable in practice. Activities will include a literature review, implementation with simulation and real data, developing map contents and format, and generalizing the method for arbitrary maps. The position requires relevant coursework, programming skills in Keras/TensorFlow, and experience with AI algorithms like neural networks.
Master thesis proposal • Highest grades from relevant Master program
3D Localization from dense RGB-D data • Knowledge of image analysis
and dedicated maps • Prociency of programming in Keras on Tensorow In the trace of upcoming hi-resolution range sensors, global • Deep understanding and experience of articial intelli- localization from dense point cloud information has emerged as a viable technique. Due to the vast amount of data, tra- gence algorithms, e.g. GA, RNN and PSO ditional deterministic techniques, such as the various types • Prociency in technical report writing of Iterative Closest Point (ICP), are unfeasible as they tend not to scale well with increased data ows and frame rates. The work will be done at CPAC Systems AB, with start In a wider sense, the before mentioned algorithms are also in the beginning of 2020 and with the duration of ≥ 20 prone to divergence should the rst initial pose lie too far weeks (30 ECTS). from the correct one, which also make them sensitive to ego velocity and a sucient, usually both unrealistic and Supervisor will be Robert Brenick, CPAC Systems unnecessary, frame rate. AB and examiner Peter Forsberg, Applied Articial In- Historically most attempts within 3D localization have telligence Group, M2, Chalmers University of Technology. made use of a point cloud (PC). Nowadays, however, there Please observe that successful candidates have to pass are many sensors producing both a PC and a color picture, CPACs as well as Applied Articial Intelligence Group [3] i.e. a RBG-D image. The advantage in making use of both criteria. the PC and the image is that the increased amount of data can give a more accurate and robust localization, given ad- If you nd this proposal interesting please send your equate processing capability for a realistic frame rate. Even application with CV and current transcript to: so, most current state-of-the-art methods use deterministic methods [1]. Robert Brenick - CPAC Systems AB Previous master thesis on the subject showed the poten- mail: Robert.Brenick@cpacsystems.se tial of the method [2]. References [1] M. Paton and J. Kosecka: Adaptive RGB-D Localiza- tion, Canadian Conference on Computer and Robot Vi- sion (CRV), June 2012. [2] S. Bastås, R Brenick: Outdoor global pose estimation from RGB and 3D data, Masters thesis - Department of Mechanics and Maritime Sciences, 2019. [3] Applied Articial Intelligence Group master thesis policy, www.me.chalmers.se/~mwahde/MasterTheses. html
Figure 1: High density point cloud [2]
This work aims at exploring new ways to make fast high
density RGB-D localization from dedicated maps and thus make the method more general and applicable in reality. Activities included in the thesis work: • Literature study of the high-density RGB-D localiza- tion problem • Implementation with simulation and real world data • Construction of suitable map contents and format • Generalization for use on arbitrary map with sucient information We are looking for two persons with a zest for AI who meet the following requirements: 1