Digital Signal Processing Reference
In-Depth Information
algorithms [1]. They have divided these algorithms into two categories: local and
global approach. Affected by various factors such as illumination, noise, shading,
lacking of texture, repetitive pattern and so on, stereo matching is usually ambiguous.
In order to remove ambiguity and obtain correct disparity, local approaches compute
the support weight in a local window to increase the signal-to-noise ratio. But if the
spanning depth of the local window is discontinuous, these approaches will lead the
foreground fattening. Using global approaches, the modeling of the stereo matching is
set to an energy function minimum problem and the disparity map is obtained by
optimization algorithm. Because of the shortcoming of the modeling, the energy of
the real disparity map is not the minimum of the energy function, so the energy of the
disparity map solved by the global approach is lower than the energy of the real
disparity map. Meanwhile, it is a NP-hard problem to search the minimum of the
function itself, and the time complexity is high [2].
Although the precision of local approaches is not as good as global, the speed is
faster. Thus these approaches are widely used in the real time systems which not
require high precision [3]. In general, the stereo matching usually includes four steps:
matching cost computation, cost aggregation, disparity computation and disparity
refinement. Local approaches only include first three steps and the research is focused
on the cost aggregation [4]. Researchers have proposed many methods from the two
dimension aggregation to three dimension aggregation such as shift-table window
[1,5], adaptive window [6,7,8,9,10,11] and adaptive support-weight window
[12,13,14,15,16,17,18,19,20]. Shift-table window method makes use of many
different windows to improve the precision of the matching. These windows are
obtained by placing the unknown matching pixel on different positions of a fixed
window. This method evaluates the cost of different windows and chooses the
window of the smallest cost as the support-window of the unknown matching pixel.
Different from shift-table window, adaptive window improve the matching's
precision by the way of changing the size of window. The method based on the
adaptive support-weight window computes different weight for every pixel in the
support-window of the unknown matching pixel, and according to the weight ratio
aggregate the matching cost of every pixel to final cost.
In this paper, we propose a new stereo matching method based on adaptive
support-weight window. First, we use the truncated absolute differences cost function
to compute the disparity space image (DSI). Second, we redefine the support-weight
of the local window which is different from the idea of the papers [13, 15 and 17]. In
a local window, the weight of every pixel is evaluated according to two factors such
as color difference and space distance between it with center pixel, and affection
produced by these factors are not independent. Finally, we aggregate the matching
cost based on the new defined weight and use the winner-take-all method to compute
the final disparity map. At the same time, we use an efficient support-weight
calculation way in order to improve method's efficiency. The results of the
experiment show that the accurate disparity map is evaluated with this method.
This paper is organized as follows. Section 2 introduces the principle of the
method. Section 3 shows the results of the method's experiment. Section 4 gives
conclusion and discussion.
