Methods for Reliable Robot Vision with a Dioptric System

Abstract

There is a growing interest in Robotics research on building robots which behave and even look like human beings. Thus, from industrial robots, which act in a restricted, controlled, well-known environment, today’s robot development is conducted as to emulate alive beings in their natural environment, that is, a real environment which might be dynamic and unknown. In the case of mimicking human being behaviors, a key issue is how to perform manipulation tasks, such as picking up or carrying objects. However, as this kind of actions implies an interaction with the environment, they should be performed in such a way that the safety of all elements present in the robot workspace at each time is guaranteed, especially when they are human beings. Although some devices have been developed to avoid collisions, such as, for instance, cages, laser fencing or visual acoustic signals, they considerably restrict the system autonomy and flexibility. Thus, with the aim of avoiding those constrains, a robot-embedded sensor might be suitable for our goal. Among the available ones, cameras are a good alternative since they are an important source of information. On the one hand, they allow a robot system to identify interest objects, that is, objects it must interact with. On the other hand, in a human-populated, everyday environment, from visual input, it is possible to build an environment representation from which a collision-free path might be generated. Nevertheless, it is not straightforward to successfully deal with this safety issue by using traditional cameras due to its limited field of view. That constrain could not be removed by combining several images captured by rotating a camera or strategically positioning a set of them, since it is necessary to establish any feature correspondence between many images at any time. This processing entails a high computational cost which makes them fail for real-time tasks. Despite combining mirrors with conventional imaging systems, known as catadioptric sensors (Svoboda et al., 1998; Wei et al., 1998; Baker & Nayar, 1999) might be an effective solution, these devices unfortunately exhibit a dead area in the centre of the image that can be an important drawback in some applications. For that reason, a dioptric system is proposed. Dioptric systems, also called fisheye cameras, are systems which combine a fisheye lens with a conventional camera (Baker & Nayar, 1998; Wood, 2006). Thus, a conventional lens is changed by one of these lenses that has a short focal length that allows cameras to see objects in an hemisphere. Although fisheye devices present several advantages over catadioptric sensors 2