6DoF State Estimation with a Mesh Constrained Particle Filter For Wheeled Robots

Abstract:
In this work, we present a highly accurate Mesh Constrained Particle Filter (MCPF) for wheeled robots. We demonstrate that the MCPF is capable of estimating poses with 6DoF in real-time on an embedded computer due to low particle count requirements. To achieve this, MCPF’s transition function constrains particle movement to a mesh surface approximating the surface the robot is traveling on. By constraining the particles, we reduce the dimensions of the effective work space the robot is operating in. In other words, the robot is effectively lying on a manifold (locally) with 3DoF embedded in SE(3). Importantly, by reducing this effective work space, significantly improved accuracy is also achieved while maintaining low particle density when compared to a dense SPF. In addition to showing improved accuracy and real-time performance, we demonstrate that the MCPF provides high levels of robustness to lost or dropped anchor measurements. Moreover, this approach has been tested on the walls of realworld storage tanks using a magnetic-wheeled crawler in the field.