Carvalho, João Luiz Carneiro; 0000-0002-6837-7838; http://lattes.cnpq.br/9595437898908532
Resumo:
Mobile robot localization is a complex task, specially in unstructured indoor environments, due to measurement noises and wrong scan-to-map association. Therefore, the quantification of uncertainty constitutes a important part of localization methods. The localization procedure becomes critical when the vehicle has low confidence about its last pose estimate, situation that requires a global localization procedure. An intuitive approach to solve the Global Localization Problem (GLP) is to distribute several pose hypotheses all over the map and select the most likely one according to an optimization heuristic such as Monte Carlo, Swarm Intelligence or Evolutionary Algorithm. However, hardware limitations and environment characteristics may affect the localization efficacy. In addition, the recent literature has few studies exploring the effectiveness and computing cost of different location methods under distinct scenarios, such as offices, corridors and large warehouses, for example. In this context, this work proposes two contributions to the Perfect Match (PM) localization algorithm: improvement of the uncertainty estimation about the pose and incorporation of the GLP. PM is a pose tracking algorithm that uses the scan-to-map maching approach and stands out for its cost-effectiveness, as it presents high accuracy and low computational cost. However, due to the kind of the algorithm, the global localization does not perform as well as the pose tracking. Furthermore, the estimation of the pose uncertainty could be improved, since it is based only on map features. The magnitude of the matching error, relevant information to indicate the quality of the estimated pose, is not taken into account by the PM implementations available in the literature. Therefore, the results presented in this work show that, in the selected scenarios, the quantification of the uncertainty about the pose by the proposed method suggests to be more adequate than the PM in its original form. Regarding the GLP, different optimization heuristics based on Evolutionary Algorithms and Swarm Intelligence were used collaboratively with the PM, such as: Particle Swarm Optimization (PSO), Differential Evolution (DE) e Genetic Algoritm (GA). Using simulations and real experiments, success rate and computing cost using different population sizes were measured. Results show that the proposed methods present different performances for different scenarios, but those based on Genetic Algorithm and Particle Swarm Optimization presented an average success rate above 83%, while other methods did not reach 80%.