Runtime Cross-Layer Optimization for Visual-Inertial Localization on Resource-Constrained Devices
Examiner Karl H. Johansson
Time: Wed 2021-06-23 09.00 - 09.30
Location: Location: Online over Zoom link: https://kth-se.zoom.us/j/4014872398
Respondent: Jessica Ivy Kelly , Reglerteknik - DCS
Opponent: Simone Morettini
Supervisor: José Araújo, Apostolos Rikos
Abstract: An increasing number of complex applications are being executed on resource constrained devices, such as drones and rovers. Such systems often operate in dynamic and unknown environments, and consequently have dynamic performance requirements based on their surroundings. These systems must consider the trade-off between application and platform performance in order to operate within resource means. This thesis proposes a runtime resource management system for a monolithic localization application. The proposed strategy uses gradient boosting regressors to predict localization accuracy and power consumption at runtime for a set of configurable application and platform parameters. A model-based controller selects parameters at runtime to optimize localization accuracy subject to a power constraint. The test-bed used for experiments consists of maplab, a visual-inertial localization and mapping framework, executed monolithically on the Nvdia Jetson AGX platform. The results highlight the importance of incorporating dynamic parameters when identifying predictive models for localization systems. The proposed system is able to track a power reference while maintaining reasonable localization accuracy at runtime, for both platform and application parameters. The results demonstrate that runtime control can achieve better performance than alternative solutions which rely on offline profiling of the configuration space.