An Adaptive Nested Particle Filter with Advanced Weighting to handle localization and tracking under extreme sensory occlusion

Robot localization under extreme occlusion is a common scenario in the real world. Examples include an autonomous car following another car along long stretches of highways, robot assistants following a human around, hospitality robots, robot pets, etc. The challenge here is that the robot's sensors get occluded to a large degree and for substantial amounts of time by the leader that the robot is trying to follow. In this scenario, localization using a usual MCL particle filter gets worse over time since observations consistently keep deviating from static expectations. In this thesis I present an Advanced Weighting approach for the Nested Particle Filter, to maintain localization under extreme occlusion while simultaneously tracking the leader. I also adapt KLD sampling to Nested MCL, which makes this approach highly scalable by allowing for dynamic variation in particle counts at both levels of nested particles, instead of ad-hoc static counts.