Fast Visual Mapping for Mobile Robot

| In this paper we describe a vision-based mapping system for mobile robot navigation that is fast, simple and requires very modest computational resources. The system comes in two parts: a visual range subsystem and a map building subsystem. The visual range subsystem acts as a kind of`virtual range sensor': it takes as input a stream of images and produces as output a stream of range-and-bearing measurements , comparable to that produced by a scanning laser range nder. The map building subsystem then fuses this data stream into a coherent grid-based representation of the environment. The system has been implemented and extensively tested on Robot J. Edgar, a small mobile robot that roams the corridors of an unmodiied ooce building. One of the key diiculties facing any mobile robot is obtaining accurate, relevant and timely maps of the environment it inhabits. Over the years, a great many sensing modalities have been applied to this task, including vision-based systems such as geometric modelling, stere-opsis and structured lighting, and non-vision systems such as sonar, radar and time-of-ight laser range-nders 1]. Unfortunately, vision-based systems have generally been distinguished by low speed and enormous computational requirements, whilst non-vision systems have been distinguished either by poor resolution (in the case of sonar) or high cost (in the case of radar and laser range-nders). In this paper, we describe a vision-based mapping system that is fast, simple and requires relatively modest computational resources. The system comes in two parts: a visual range subsystem and a map building subsystem. The visual range subsystem acts as a kind of`virtual range sen-sor', which takes as input a stream of images and produces as output a stream of range-and-bearing measurements. The map building subsystem then fuses this data stream into a coherent grid-based representation of the environment , known as anòccupancy map' 2]. The motivation for this design arises from the observation that very good environment maps can be obtained by combining scanning laser range nders with grid-based data fusion techniques 3]. In our system, the laser range nder is replaced with a vision-based virtual range sensor that produces comparable output. The system uses standard vision components { a camera, frame grabber and computer { and as such is low cost. For example, the results presented in this paper where obtained using a inexpensive monochrome camera tted with a plastic lens, a 10-year-old frame grabber and 33MHz 486-based computer. In …