Using image signatures for place recognition

Abstract

This paper describes and evaluates an image-based place recognition method for mobile robots based on matching image signatures. The method represents an image as a set of arrays of coarse-scale measurements over subimages. We describe several different measurement functions we devised, and show how they can be used conjunctively to improve recognition accuracy. Experimental results show that this method enables accurate place recognition, comparable to recent results for localization using occupancy grids, and also to recent results for image-based fingerprint recognition.

Introduction

A central problem for a mobile robot is localization, determining its location from sensed data. When a complete model of the environment is not available a priori, this problem may become quite difficult. In the topological mapping paradigm (Kuipers and Byun, 1988; Levitt et al., 1987), the environment is discretized into a number of different `places', so that the world can be represented as a graph of these places, whose arcs represents possible movements between places. This makes localization easier, as the problem reduces to recognizing at which place, of a given set, the robot is located. A number of methods have been proposed for place recognition, including certainty grids (Elfes, 1987, Yamauchi and Langley, 1997) and local geometry (from vision) (Braunegg, 1990, Taylor and Kriegman, 1992). Both of these types of methods are computationally intensive and can be sensitive to even small changes in the environment. In this paper we empirically evaluate another approach, that of image matching, in which recognition is performed by matching 2D representations of input images. The method we use is based on image signatures, coarse-scale representations of image properties such as edge-direction or image homogeneity. An image signature is an array of values, each derived by a measurement of an image property over a portion of the input image (as depicted in Fig. 1). Computing and matching image signatures is efficient, and is insensitive to small changes in the environment.

Similar image matching approaches have previously been applied to robotic and other recognition tasks. For example, Horswill (1993) used coarse grey-scale images for place recognition for a mobile robot, using strong geometric constraints. 2D image matching has also been used extensively for non-robotic recognition tasks. In particular, grey-scale correlation of image templates has been used for face recognition (Brunelli and Poggio, 1993), iris recognition (Wildes et al., 1996), and finding volcanoes in images of Venus (Wiles and Forshaw, 1993). Grey-scale correlation works well in such cases, where well-defined templates (models of image regions) exist.

A different set of techniques are the feature-based approaches, in which a set of feature values are extracted from an input image and the resultant feature vector are matched to a set of model feature vectors. Feature-based recognition has been applied to 2D object recognition (Crowley and Sanderson, 1987, Sethi and Ramesh, 1992, Beveridge and Riseman, 1997). Similar approaches have also been applied to the problem of image retrieval from multimedia databases (Jin et al., 1996, Manjunath and Ma, 1996, Schmid and Mohr, 1997). Color histograms have also been used successfully for image retrieval (Swain, 1990, Mehtre et al., 1995).

Our method of using image signatures incorporates some of the benefits of both image-based and feature-based approaches. The approach is based on Nelson's (Nelson, 1989) use of directional image signatures for stimulus-response robotic navigation. We extend this to include signatures for several different types of image measurements, including edge direction and strength, as well as texture-based measures. By using feature-based measures, rather than straight grey-scale values, we expect to attain more robust recognition in the face of illumination or small environmental changes. On the other hand, by looking at measures over large subimages, rather than individual features, we expect recognition to be insensitive to larger environmental changes, such as moved objects and the like.

An approach similar to that described in this paper was used for fingerprint recognition by Coetzee and Botha (1993). There, the authors used a type of image signature computed from a skeletonized image of a fingerprint, and applied a variety of matching techniques for recognizing fingerprints. The task of place recognition differs from fingerprint recognition, however, in that images of the world are more visually complex than fingerprints, while on the other hand the differences between fingerprints may be very subtle. In addition, the place recognition method described in this paper also explicitly considers the possibility of changes in viewer position at the same location.

This paper is organized as follows. In the following section, we will describe how image signatures are computed, including a description of several measurement functions that we use, and how signatures are compared for similarity. We also describe a conjunctive matching scheme, where several measurement functions are used in tandem for recognition. We develop measures for evaluating the quality of different measurement functions and combinations thereof when developing a recognition system based on image signatures. Section 2describes a series of experiments that demonstrate how the method of image signatures is useful for place recognition. We conclude that the method of image signatures is useful for matching complex images, particularly when using several measurement functions together.