Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
High Performance Computation for Non-Linear Estimation
Analyze and optimize a novel non-linear optimization framework for low-latency visual-inertial state estimation.
Keywords: visual inertial odometry state estimation non-linear nonlinear optimization backend low latency high performance computing embedded systems
Visual-inertial odometry (VIO) has matured and became the go-to solution for mobile robot state estimation.
There exists a variety of VIO pipelines, reaching from computationally-efficient filter-based approaches to more accurate optimization-based sliding window estimators.
However, such sliding window estimators are often relatively slow and introduce significant latency.
We are developing a novel optimization-based backend with focus on low latency estimation.
This thesis will focus on improving the computational efficiency of our new backend by analyzing the existing framework, reporting bottlenecks, and implementing new optimization strategies or exploiting existing acceleration libraries (such as BLAS, LAPACK or higher level libraries).
There are multiple target platforms, including small ARM single-board computers on drones, on which the final solution will be demonstrated.
Strong c++ knowledge is needed and experience in numerical optimization is welcome.
While there is a high focus on code development, there are options for theoretical contributions as well.
Visual-inertial odometry (VIO) has matured and became the go-to solution for mobile robot state estimation. There exists a variety of VIO pipelines, reaching from computationally-efficient filter-based approaches to more accurate optimization-based sliding window estimators. However, such sliding window estimators are often relatively slow and introduce significant latency. We are developing a novel optimization-based backend with focus on low latency estimation. This thesis will focus on improving the computational efficiency of our new backend by analyzing the existing framework, reporting bottlenecks, and implementing new optimization strategies or exploiting existing acceleration libraries (such as BLAS, LAPACK or higher level libraries). There are multiple target platforms, including small ARM single-board computers on drones, on which the final solution will be demonstrated. Strong c++ knowledge is needed and experience in numerical optimization is welcome. While there is a high focus on code development, there are options for theoretical contributions as well.
The goal of this thesis is to analyze and profile an existing optimization backend for visual-inertial state estimation and propose, implement, and benchmark improvements for low-latency execution on handheld devices and drones.
These improvements can be code optmiziations, combination with existing libraries for sparse and accelerated linear algebra, or theoretical advances.
The goal of this thesis is to analyze and profile an existing optimization backend for visual-inertial state estimation and propose, implement, and benchmark improvements for low-latency execution on handheld devices and drones. These improvements can be code optmiziations, combination with existing libraries for sparse and accelerated linear algebra, or theoretical advances.