Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Learning Redundant Information for Graph-Based SLAM
The goal of this project is to investigate how machine learning techniques can be used to identify redundant data in a SLAM graph, which can then be removed to boost the efficiency and scalability of SLAM.
State-of-the-art Simultaneous Localization And Mapping (SLAM) systems have reached a high level of maturity in single-agent [1] and multi-agent [2] scenarios. However, while showing substantial accuracy and robustness in various environments, the scalability of these solutions (in terms of participating agents and size of the workspace) remains a problem, since the amount of data that can be efficiently handled by the SLAM system is limited. One approach to mitigate this problem is to employ redundancy detection scheme to sparsify the SLAM graph, since not every chunk of data in the graph contributes the same amount of information (such as data from repeated visits of the same location). As shown in [3], redundancy detection can significantly compress SLAM graphs with minimal loss in accuracy. However, while [3] is based on information-theoretic notions, the goal of this project is to investigate how machine learning techniques can be used for this purpose instead, and whether they can achieve a better performance than traditional theoretic and heuristics approaches currently present in the literature, in order to further boost the scalability of contemporary SLAM solutions.
References:
[1] Qin et al.: VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator, T-RO 2018
[2] Karrer et al.: CVI-SLAM – Collaborative Visual-Inertial SLAM, RA-L 2018
[3] Schmuck and Chli: On the redundancy detection in keyframe-based SLAM, 3DV 2019
State-of-the-art Simultaneous Localization And Mapping (SLAM) systems have reached a high level of maturity in single-agent [1] and multi-agent [2] scenarios. However, while showing substantial accuracy and robustness in various environments, the scalability of these solutions (in terms of participating agents and size of the workspace) remains a problem, since the amount of data that can be efficiently handled by the SLAM system is limited. One approach to mitigate this problem is to employ redundancy detection scheme to sparsify the SLAM graph, since not every chunk of data in the graph contributes the same amount of information (such as data from repeated visits of the same location). As shown in [3], redundancy detection can significantly compress SLAM graphs with minimal loss in accuracy. However, while [3] is based on information-theoretic notions, the goal of this project is to investigate how machine learning techniques can be used for this purpose instead, and whether they can achieve a better performance than traditional theoretic and heuristics approaches currently present in the literature, in order to further boost the scalability of contemporary SLAM solutions.
References:
[1] Qin et al.: VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator, T-RO 2018