Fast Line-Segment Extraction for Semi-dense Stereo Matching
This paper describes our work on practical stereo vision for mobile robots using commodity hardware. The approach described in this paper is based on line segments, since those provide a lot of information about the environment, provide more depth informa
- PDF / 2,247,206 Bytes
- 13 Pages / 430 x 660 pts Page_size
- 1 Downloads / 225 Views
Abstract. This paper describes our work on practical stereo vision for mobile robots using commodity hardware. The approach described in this paper is based on line segments, since those provide a lot of information about the environment, provide more depth information than point features, and are robust to image noise and colour variations. However, stereo matching with line segments is a difficult problem due to poorly localized end points and perspective distortion. Our algorithm uses integral images and Haar features for line segment extraction. Dynamic programming is used in the line segment matching phase. The resulting line segments track accurately from one frame to the next, even in the presence of noise.
1
Introduction
This paper describes our approach to stereo vision processing for urban search and rescue (USAR) robotics using low-quality cameras. The goal of all stereo vision processing is to produce dense depth maps that can be used to accurately reconstruct a 3-D environment. However, our target applications impose additional constraints: (a) since the stereo system is mounted on a mobile robot, the approach must support real-time depth map extraction, and (b) since we are using cheap off-the-shelf webcams, the approach must be robust to noise and other errors. The two algorithms described in this paper include fast line segment extraction and two frame line segment matching. The fast line segment extraction makes use of integral images to improve the performance of gradient extraction, as well as provide an edge representation suitable for binary image segmentation. Line segment matching is achieved using a dynamic programming algorithm, with integral images providing fast feature vector extraction. Most of the focus in stereo vision processing has been on point featurebased matching. The most notable feature extraction methods are corner detection [1,2], SIFT [3], and SURF [4]. Most point feature based stereo matching relies on the epipolar constraint [5], to reduce the search space of the feature matcher to a 1-D line. The epipolar constraint can be applied using the Essential or Fundamental matrix, which can be determined using several methods [5]. G. Sommer and R. Klette (Eds.): RobVis 2008, LNCS 4931, pp. 59–71, 2008. c Springer-Verlag Berlin Heidelberg 2008
60
B. McKinnon and J. Baltes
The problem that was encountered early in our stereo vision research was that the cameras tended to shake while the robot would move. Even small changes in the position of one camera changes relative to a second require the recalculation of the epipolar lines, which can be an expensive and error-prone task during the operation of the robot. This problem demanded a more robust feature set, that would be trackable over a larger 2-D search space. In earlier work[6], our approach focused on region-based stereo matching. Region segmentation proved to be a difficult task in complex scenes under realtime constraints. In addition, regions would not always form in a similar manner, making our hull signature matching
Data Loading...
