An Improved Observation Model for Monte-Carlo Localization Integrated with Reliable Reflector Prediction

Abstract

Robust and reliable localization is a fundamental prerequisite for many applications of mobile robots which can be achieved in many ways. Among them, landmark-based localization is a proven and practical technique. However, incorrect detection of landmarks can seriously affect the positioning result. This paper presents an improved observation model for Monte-Carlo localization (MCL), which improves the robustness of localization by reliable reflector prediction in the ambiguous environments caused by incorrect reflectors detection. This improved observation model is based on a Bayesian network into which the reflective intensity of LIDAR and the labeled grid-map are considered. And then, a reflector probability field and a learned intensity observation model are proposed to accomplish fast probabilistic inference of robot pose based on our improved observation mode. Also, a variant MCL, called reflector prediction-based Monte-Carlo localization (RP-MCL) is realized based on our observation model. The effectiveness of the RP-MCL is verified in real-world scenarios using our self-developed robot and the results demonstrate that our observation model improves the localization performance in the ambiguous environments.