Articulated pose estimation via over-parametrization and noise projection

Abstract

Outside the factory, robots will often encounter mechanical systems with which they need to interact. The robot may need to open and unload a kitchen dishwasher or move around heavy construction equipment. Many of the mechanical systems encountered can be described as a series of rigid segments connected by joints. The pose of a segment places constraints on adjacent segments because they are mechanically 'connected. When modeling or perceiving the motion of such an articulated system, it is beneficial to make use of these constraints to reduce uncertainty. In this thesis, we examine two aspects of perception related to articulated structures. First, we examine the special case of a single segment and recover the rigid body transformation between two sensors mounted on it. Second, we consider the task of tracking the configuration of a multi-segment structure, given some knowledge of its kinematics. First, we develop an algorithm to recover the rigid body transformation, or extrinsic calibration, between two sensors on a link of a mobile robot. The single link, a degenerate articulated object, is often encountered in practice. The algorithm requires only a set of sensor observations made as the robot moves along a suitable path. Over-parametrization of poses avoids degeneracies and the corresponding Lie algebra enables noise projection to and from the over-parametrized space. We demonstrate and validate the end-to-end calibration procedure, achieving Cramer-Rao Lower Bounds. The parameters are accurate to millimeters and milliradians in the case of planar LIDARs data and about 1 cm and 1 degree for 6-DOF RGB-D cameras. Second, we develop a particle filter to track an articulated object. Unlike most previous work, the algorithm accepts a kinematic description as input and is not specific to a particular object. A potentially incomplete series of observations of the object's links are used to form an on-line estimate of the object's configuration (i.e., the pose of one link and the joint positions). The particle filter does not require a reliable state transition model, since observations are incorporated during particle proposal. Noise is modeled in the observation space, an over-parametrization of the state space, reducing the dependency on the kinematic description. We compare our method to several alternative implementations and demonstrate lower tracking error for fixed observation noise.