Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Adaptive Trajectory Refinement for SLAM
The aim of this project is to develop effective means and algorithms to detect over- or underparameterized trajectory estimates in order to reduce the state size in the underlying non-linear least squares Simultaneous Localization and Mapping (SLAM) problem. Specifically, the alluded adaptive refine
Keywords: Continuous-time Simultaneous Localization and Mapping, Adaptive Trajectory Refinement
SLAM is the task of maneuvering in an, a priori, unknown space while mapping its environment and simultaneously estimating the relative transformation of the moving reference frame in said map. Within the context of mobile robotics, gaining an accurate and useful understanding of a robot's surroundings is elementary to successfully achieve any autonomous tasks, such as navigation or obstacle avoidance. Despite aforementioned task representing one of the most well-studied problems in robotics, the question of how to compactify the pose estimate and the landmark map on-the-fly without detrimentally influencing the overall accuracy of the pipeline is still subject to ongoing research. Within the context of standard SLAM algorithms, one often relies on the formalism of key-frames to limit bloating the optimization state size. In particular, one selects a subset of frames, which ought to be representative for the trajectory as a whole. As previously alluded, selecting aforementioned key-frames is non-trivial if loss of accuracy should be avoided. Besides the conventional approaches, another approach, based on representing pose estimates in continuous-time, has gained traction in the scientific community within recent years. Its most common formulation, however, does not allow for modifications to be made after initialization. Owing to the fact that one can not refine or coarsen the underlying trajectory, a system based on it is bound to not make optimal use of the available optimization time due to a wastefully large state size, or not capture the apparent motion correctly due to losses in details of the pose estimates. Thus, we aspires to explore intelligent ways of combining a continuous-time representation along the possibility of dynamically refining state estimates where required.
SLAM is the task of maneuvering in an, a priori, unknown space while mapping its environment and simultaneously estimating the relative transformation of the moving reference frame in said map. Within the context of mobile robotics, gaining an accurate and useful understanding of a robot's surroundings is elementary to successfully achieve any autonomous tasks, such as navigation or obstacle avoidance. Despite aforementioned task representing one of the most well-studied problems in robotics, the question of how to compactify the pose estimate and the landmark map on-the-fly without detrimentally influencing the overall accuracy of the pipeline is still subject to ongoing research. Within the context of standard SLAM algorithms, one often relies on the formalism of key-frames to limit bloating the optimization state size. In particular, one selects a subset of frames, which ought to be representative for the trajectory as a whole. As previously alluded, selecting aforementioned key-frames is non-trivial if loss of accuracy should be avoided. Besides the conventional approaches, another approach, based on representing pose estimates in continuous-time, has gained traction in the scientific community within recent years. Its most common formulation, however, does not allow for modifications to be made after initialization. Owing to the fact that one can not refine or coarsen the underlying trajectory, a system based on it is bound to not make optimal use of the available optimization time due to a wastefully large state size, or not capture the apparent motion correctly due to losses in details of the pose estimates. Thus, we aspires to explore intelligent ways of combining a continuous-time representation along the possibility of dynamically refining state estimates where required.
- Literature review and familiarization with SLAM pipelines.
- Conceptualization and prototyping of insertion and removal strategies/algorithms.
- Comparison against existing approaches.
- (Optional) Integration into and testing within a larger SLAM framework in C++.
- Literature review and familiarization with SLAM pipelines. - Conceptualization and prototyping of insertion and removal strategies/algorithms. - Comparison against existing approaches. - (Optional) Integration into and testing within a larger SLAM framework in C++.
- Solid mathematical background and maturity with Matlab and/or C/C++.
- Prior knowledge about Simultaneous Localization and Mapping is beneficial.
- Some familiarity with Linux, ROS, Eigen and Ceres is desirable.
- Solid mathematical background and maturity with Matlab and/or C/C++. - Prior knowledge about Simultaneous Localization and Mapping is beneficial. - Some familiarity with Linux, ROS, Eigen and Ceres is desirable.