Automatic Refinement of Foreground Regions for Robot Trail Following Mehmet Kemal Kocamaz & Christopher Rasmussen University of Delaware kocamaz@udel.edu, cer@cis.udel.edu Abstract Continuous trails are extended regions along the ground such as roads, hiking paths, rivers, and pipelines which can be navigationally useful for ground-based or aerial robots. Finding trails in an im- age and determining possible obstacles on them are im- portant tasks for robot navigation systems. Assuming that a rough initial segmentation or outline of the region of interest is available, our goal is to refine the initial guess to obtain a more accurate and detail representa- tion of the true trail borders. In this paper, we compare the suitability of several previously published segmenta- tion algorithms both in terms of agreement with ground truth and speed on a range of trail images with diverse appearance characteristics. These algorithms include generic graph cut, a shape-based version of graph cut which employs a distance penalty, GrabCut, and an it- erative superpixel grouping method. 1 INTRODUCTION Vision-based trail finding and tracking can be con- sidered as a form of road following [7, 13, 12, 5]. How- ever, several factors make the computer vision task par- ticularly hard, including indistinct borders, abrupt ele- vation changes, dead-ends and forks, sharply varying illumination conditions due to shadows, a wide range of trail materials and hence colors and textures, and the possibility of in-trail objects such as rocks, stumps, or grass. In our previous work [8], we described an effi- cient and robust approach to trail finding using a general model of color contrast and a triangular shape template. A shortcoming of the approach, however, was the ap- proximation of the trail shape in its linearity and lack of local shape variation. In this paper we describe a second stage of shape es- timation in which the initial, rough shape is refined au- tomatically, without user interaction. The contribution of this work is a survey, analysis, and comparison of several basic segmentation algorithms for this purpose, specifically graph cut [2], graph cut with distance maps [6], GrabCut [10], and a grouping method based on su- perpixel oversegmentation [4]. Although the focus here is on trail image data most relevant to mobile robot ap- plications, we believe that the problem of automatically refining segmentations is a general one. The rest of the paper is organized as follows. Sec- tion 2 outlines the basic segmentation methods. Sec- tion 3 describes some algorithmic changes we made to transform the GraphCut methods and GrabCut into automatic foreground refinement algorithms. Section 4 shows some results of the transformed methods on our data sets, compares their performances and the final segmentations of the techniques with ground-truth data. Finally, in Section 5, we summarize the algorithms and their results. 2 Review of Basic Segmentation Algo- rithms Graph Cut: The segmentation algorithm proposed in [3, 1, 2] can convert an image segmentation task to a foreground/background labeling problem. Labeling problem is to assign a label L i for each pixel i in the image. L i is from the set of segmentation result, S result = (”background”, ”object”). L = (L 1 ,L 2 , ..., L n ), where n is the number of the pixels in the image, is the solution of the segmentation. To find the set L the following energy function can be solved: E(L)= n  i=0 R i (L i )+ λ  {i,j}∈N B(L i ,L j ) (1) In this energy function, R i (L i ) is called as regional term and B(L i ,L j ) as boundary term. λ ≥ 0 is a weight term to set the relative influence of boundary term versus regional term in the function. N is all un- ordered neigborhood pixel pairs. 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.991 4061 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.991 4081 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.991 4077 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.991 4077 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.991 4077