Onboard perception and localization for resource-constrained dynamic environments : a RoboCup small size League Case Study

• Autor(a) principal: MELO, João Guilherme Oliveira Carvalho de
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional da UFPE
• Texto Completo: https://repositorio.ufpe.br/handle/123456789/52045
• topic: Engenharia da computação; Robôs móveis autônomos; Auto-localização; Monte Carlo Localization; RoboCup

Self-localization consists of estimating a robot’s position and orientation (pose) regarding its operating environment and is a fundamental skill in autonomous mobile robot navigation. Monte Carlo Localization (MCL) is a particle filter-based algorithm that addresses the local- ization problem by maintaining a set of particles that represent multiple hypothesis of the current robot’s state. At each step, the particles’ are moved according to the robot’s motion and their likelihoods, also called importance weights, are estimated based on the similarities between measurements acquired by the robot and their expected values, given the particle state. Then, a resampling algorithm is applied to the distribution, generating a new set based on the current weights. MCL finds successful utilization in RoboCup robot soccer leagues for solving the self- localization problem in humanoid and standard platform competitions. At 2022, this problem was also introduced in the RoboCup Small Size League (SSL) within the Vision Blackout Technical Challenge, which restricts teams to use onboard sensing and processing only for executing basic soccer tasks, instead of the typical SSL approach that uses an external camera for sensing the environment, but no solutions were proposed for self-localization so far. There- fore, this work presents an integrated pipeline for solving the SSL self-localization problem while also detecting the environment’s dynamic objects, using onboard monocular vision and inertial odometry data. We enhance the MCL using insights from implementations of other RoboCup leagues, im- proving the algorithm’s robustness regarding imprecise measurements and motion estimations. Also, we increase the algorithm’s processing speed by adapting the number of particles in the set according to the confidence of the current distribution, also called Adaptive MCL (AMCL). For that, we propose a novel approach for measuring the quality of the current distribution, based on applying the observation model to the resulting particle of the algorithm. The ap- proach was able to drastically increase the system’s computation speed, while also maintaining the capability to track the robot’s pose, and the confidence measure may be useful for making decisions and performing movements based on the current localization confidence.