Sensor planning for object pose estimation and identification

Abstract

This paper proposes a novel approach to sensor planning for simultaneous object identification and 3D pose estimation. We consider the problem of determining the next-best-view for a movable sensor (or an autonomous agent) to identify an unknown object from among a database of known object models. We use an information theoretic approach to define a metric (based on the difference between the current and expected model entropy) that guides the selection of the optimal control action. We present a generalized algorithm that can be used in sensor planning for object identification and pose estimation. Experimental results are also presented to validate the proposed algorithm.

Dynamic sensor planning with stereo for model identification on a mobile platform

“…The use of the pan-tilt-unit (PTU) control loop is one of the additions to our previous work [1]. The independent movement of the camera on the PTU allows the robot to take more optimal paths toward a goal location without the added constraints of keeping the object in the field of view at all times while simultaneously satisfying the vehicle's nonholonomic wheel constraints.…”

“…2 which shows the key reference frames and the spatial transformations between reference frames at one timestep and the next 1 . The transformation between the object frame and camera frame follows the recursive prediction equation:…”

“…The transform G O k−1 ,O k is derived from the object's motion model which we assume to be a random walk perturbed by 1 The notation used here is that the transform G AB ∈ SE(3) transforms a 3D point defined in frame B into the reference frame of A. There are three reference frames considered: the robot reference frame (R), the camera reference frame (C), and the object reference frame (O).…”

“…Our approach is an extension of our previous work ( [1]) in which the motion of the object-to-be-identified was constrained to lie within the field of view of the stereo camera, and the motion of the robot was confined to a circular ring about the object. In this paper neither the object nor the mobile agent are constrained in their motion.…”