|Estimating the vanishing point of a sidewalk
|WSCG 2014: communication papers proceedings: 22nd International Conference in Central Europeon Computer Graphics, Visualization and Computer Visionin co-operation with EUROGRAPHICS Association, p. 369-374.
|Václav Skala - UNION Agency
|rozpoznání chodníku;detekce řádků;odhad mrtvých bodů;obrazy přírodní scény
|Keywords in different language:
|sidewalk recognition;line detection;vanishing point estimation;natural scene images
|Abstract in different language:
|In this paper, we propose an algorithm estimating the vanishing point (VP) of a sidewalk in a man-made environment from a single image. For finding the VP, the lines in an image are efficiently exploited as a clue because the projections of parallel lines in an image intersect at a VP. However, there are too many noises disturbing line detection in natural scene images, for example trees, pedestrian and shadows. Thus, we suggest a noise reduction technique called orientation consistency pass filtering (OCPF) for improving line detection performance. An edge orientation at the window center of OCPF is compared to its neighbor edge orientations for calculating orientation difference. The center pixel is removed if the difference of edge orientations greater than a threshold, and preserved otherwise. In addition, we suggest a novel vanishing point detection method using edge orientation voting (EOV), which predict VP position accurately. The lines filtered by OCPF can generate the VP candidates using bottom-up extended lines. The VP candidates receive supports from all edges below the currently inspected VP candidate. The most supported VP candidate is selected as dominant VP. This proposed method was implemented and tested on 600 sidewalk image database that has 640x320 resolutions. 74.3 % of the test sidewalk images are in range from 0 pixels to 20 pixels from manually marked VP.
|@ Václav Skala - UNION Agency
|Appears in Collections:
|WSCG 2014: Communication Papers Proceedings
Please use this identifier to cite or link to this item:
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.