On Performance portability Via Runtime Adaptation for VO/VSLAMs

  • Abdullah Khalufa

Student thesis: Phd


The development and use of Visual Simultaneous Localisation and Mapping (VSLAM) in many diverse navigation and vision applications has resulted in a wide variety of algorithmic implementations. Deploying these VSLAM implementations on energy-constrained platforms dictates balancing the target application performance objectives through the tuning of hyperparameters at runtime to ensure prolonged and robust operation. Balancing these performance objectives is difficult to achieve in a portable fashion due to the diverse settings in which different VSLAM implementations operate, leading to limited and over-fitted application-specific solutions. Further, balancing the performance objectives at runtime requires solutions that are lightweight and easy to maintain on existing and emerging VSLAM implementations. This thesis explores the idea of performance portability at a VSLAM macro-benchmark level where improvements in performance are achieved via runtime adaptation. The research focuses on three aspects: 1. The use of portable metrics for characterising motion and scene changes exploited to obtain effective guidance for runtime adaptations; 2. A top-down approach to enable coping with different VSLAM deployment environments and settings to show that this can be achieved in an efficient and portable manner; 3. The capability of the proposed framework to improve runtime performance, explored and evaluated in diverse and challenging settings where the framework is shown to improve performance objectives with minimal impact on the overall accuracy and robustness of well-established VSLAM implementations. Further, portable performance is shown to be fully achievable using the framework for formulations with similar levels of computational intensity, and is generally achievable to a useful extent.
Date of Award1 Aug 2021
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorGraham Riley (Supervisor) & Mikel Luján (Supervisor)

Cite this