CN101840574B - Depth estimation method based on edge pixel characteristics - Google Patents

Abstract

The invention discloses a depth estimation method based on edge pixel characteristics, which mainly solves the problem of inaccurate image edge depth estimation in the conventional depth estimation method in an FTV system. The scheme is as follows: firstly, dividing pixels in an image into three types of pixels on an edge, pixels beside the edge and non-edge pixels according to the specific position of a current pixel; secondly, respectively designing a corresponding parallax non-uniformity function of each type of pixels according to the depth characteristics of the edge pixels of the object in the image; then, respectively calculating pixel brightness inconsistency and parallax inconsistency according to the brightness inconsistency function and the obtained three types of parallax inconsistency functions, and performing corresponding parallax estimation by using an energy minimization function; and finally, converting the parallax value into a corresponding depth value by using a parallax depth conversion function according to the estimated parallax value to finish depth estimation. The invention effectively improves the depth estimation accuracy of the object edge pixels and can effectively ensure the subjective quality and the objective quality of the virtual view synthesized by the receiving end of the FTV system.

Description

Depth Estimation Method Based on Edge Pixel Features

technical field

The invention belongs to the communication field, and relates to depth estimation technology in three-dimensional stereoscopic video, specifically a depth estimation method capable of obtaining a high-precision depth map, so that the synthetic virtual view has fewer noise points formed at the edge part, and is effective The subjective quality and objective quality of the synthesized virtual view are improved, and it can be applied to any viewpoint television system.

Background technique

In traditional TV systems, users can only watch a limited viewing angle in the three-dimensional world, and the viewing point and viewing angle are determined by the three-dimensional space position and direction of the camera. Therefore, the user cannot freely select the viewpoint and viewing angle for viewing. Any point of view TV FTV system allows users to watch the real three-dimensional space from different perspectives, thus providing a new, more vivid and real three-dimensional audio-visual system. Since FTV can provide users with a more realistic interactive viewing effect, it can be widely used in various types of video systems such as broadcast communication, entertainment, education, medical treatment, and video surveillance.

Figure 1 shows the main functional modules of the FTV system. The data generated by the sending end of the FTV system includes the video data captured by the camera array located at multiple viewpoints, and the corresponding scene depth data; the receiving end of the FTV system uses various depth information-based virtual view generation Viewpoint video data. Therefore, the acquisition of high-quality depth data is one of the key technologies to realize the FTV system. In the current FTV system, the disparity value d is first obtained by disparity estimation, and then the disparity value d is converted into the corresponding depth value according to the following formula:

In the formula: I represents the distance between cameras; f is the focal length of the camera lens; Z _near and Z _far represent the depth value of the plane with the closest distance and the depth value of the plane with the farthest distance from the object in the 3D scene to the camera, respectively.

Currently, graph cut-based global optimization is mainly used to estimate disparity. That is, firstly, an energy function representing the non-uniformity of pixel brightness and disparity is established, and then the energy function is minimized by global optimization.

Energy minimization defines the optimal parallax combination f that can minimize the parallax non-uniform energy function E, and the energy function consists of the brightness non-uniformity item of the corresponding parallax pixel point and the parallax non-uniformity item of adjacent pixels Composition, as shown in formula (2):

E(f)＝E _data (f)+E _smooth (f) (2)

P表示当前视图中的象素集合，D_p衡量当前视图中的象素p在视差f_p的条件下与参考视图中的象素(p+f_p)之间的亮度非一致性，通常

其中I_p表示当前视图中的象素p的亮度值，

表示参考视图中的象素(p+f_p)的亮度值；而

表示当前象素与其相邻象素之间的视差非一致性，其中，N表示邻近象素对的集合，V_{p，q}∈N(f_p，f_q)衡量当前视图中象素p的视差为f_p，其相邻象素q的视差为f_q时，二者视差的不一致性。若：In the formula:

P represents the set of pixels in the current view, and D _p measures the brightness inconsistency between the pixel p in the current view and the pixel (p+f _p ) in the reference view under the condition of parallax f _p , usually

Where I _p represents the brightness value of pixel p in the current view,

represents the brightness value of the pixel (p+f _p ) in the reference view; and

Indicates the disparity inconsistency between the current pixel and its neighboring pixels, where N represents the set of neighboring pixel pairs, and V _{{p, q}∈N} (f _p , f _q ) measures the pixel p in the current view When the disparity of f _p is f p , and the disparity of its adjacent pixel q is f _q , the disparity between the two is inconsistent. like:

V _{{p, q}∈N} (f _p , f _q )=λ|f _p -f _q | (3)

Then for smaller |f _p -f _q |, the corresponding disparity inconsistency cost will not be too large; and for larger |f _p -f _q |, the disparity inconsistency cost will increase accordingly. The resulting optimal disparity is thus usually not very different from the disparity between adjacent pixels. In fact, in the stereo matching problem, due to the discontinuity of parallax, especially at the edge of the object, the parallax of the pixel at the edge of the object is usually quite different from the parallax of the surrounding pixels. Therefore, formula (3) will Causes the over-smoothing phenomenon of the edge part of the object. Due to the over-smoothing of the edge part of the object, it will cause blurring and aliasing of the edge part in the synthetic view, which seriously affects the visual effect.

For objects on the same edge, their disparity is basically the same, even if there is a change, it is continuous and slow, so formula (3) is adaptive. However, due to the projection of different objects in the image divided by the edges, they do not have the consistency of disparity even for adjacent pixels. Therefore, it is necessary to study the adaptability of formula (3) in order to produce a truly effective depth estimation method.

At present, the depth estimation method used in the FTV system is to minimize the parallax inconsistency energy function E by setting the optimal parallax f _p , and finally find the appropriate parallax value for each pixel, and then use the formula (1 ) converts the disparity value d to the corresponding depth value. The disparity inconsistency energy function E is composed of the brightness inconsistency item of the corresponding parallax pixel point and the disparity inconsistency item of adjacent pixels. Among them, the brightness inconsistency item corresponding to the disparity pixel describes that a disparity f _p is assigned to the pixel p in the current view, and the corresponding pixel point of the pixel p is found in the reference view through the disparity f _p , and the current pixel p The degree of matching with its corresponding pixel; the non-consistency item of adjacent pixel parallax describes the inconsistency of the parallax between pixel p and its surrounding adjacent pixels in the current view, if the current pixel p and surrounding adjacent pixels q has different disparities, the corresponding inconsistency cost is added in the disparity inconsistency energy function.

Paying attention to the particularity of the parallax of edge points, the existing depth estimation method considers that compared with the surrounding pixels, the edge is a pixel point with step-jump parallax. In order to study the depth of edge pixels, this method gives some edge auxiliary information of scene images, and the edge pixels pointed out by the edge auxiliary information have step disparity.

As shown in Figure 2, the existing depth estimation method first judges the state of the current pixel and its surrounding adjacent pixels, that is, judges whether the pixel is located on the edge of the object in the image, and then designs two Parallax-like inconsistency cost functions V _j and V _c :

其中，E为边缘图    (4)

Among them, E is the edge map (4)

即如果象素p和象素q均不在边缘上，它们之间的视差变化较为平缓，因此若出现较大的阶跃视差，V_c取一个无穷大值；

而若象素p和象素q其中一个位于边缘上或者二者均位于边缘上，则它们之间的视差具有较大的阶跃，因此若二者之间的视差变化较为平缓，V_j将同样取一个无穷大值。In the formula:

That is, if the pixel p and the pixel q are not on the edge, the parallax between them changes relatively gently, so if there is a large step parallax, _Vc takes an infinite value;

And if one of pixel p and pixel q is located on the edge or both are located on the edge, the disparity between them will have a large step, so if the disparity between the two is relatively gentle, V _j will be Also take an infinite value.

Secondly, calculate the pixel brightness inconsistency and calculate the parallax inconsistency according to the designed parallax inconsistency function, and use the energy minimization function to estimate the parallax.

Finally, using the estimated disparity value, the disparity value is converted into the corresponding depth value according to the disparity depth conversion formula to complete the depth estimation.

The existing depth estimation method is based on the fact that the parallax of points on the same edge should change gently; at the same time, the method points out that the parallax difference between pixels on the edge of the image and pixels outside the edge of the image is relatively large. In practice, if the edge pixel and its adjacent pixels belong to the same depth object, the disparity value of the two should change smoothly; if the two do not belong to the same depth object, the disparity difference may be large. This method indiscriminately considers that the parallax difference between the edge pixel of the image and the adjacent pixel is large, and the wrong parallax estimation result will be generated at the edge of the image, thereby affecting the accuracy of the depth value at the edge of the image, making Composite virtual views are of reduced quality.

Contents of the invention

The purpose of the present invention is to address the problems existing in the existing depth estimation methods in the FTV system, and propose a depth estimation method based on edge pixel depth features to improve the depth estimation results at the edge of the object, thereby ensuring that the receiving end of the FTV system synthesizes virtual view quality.

The technical solution for realizing the present invention is: according to the specific position of the current pixel, the pixels in the image are divided into three categories; Inconsistency function; use the obtained disparity inconsistency function to estimate the disparity; convert the disparity value into the corresponding depth value, and the specific steps include:

A. The pixel in the image is classified, and the pixel in the image is divided into three classes: the first class is the pixel on the edge, the second class is the pixel next to the edge, and the third class is the non-edge pixel;

B. According to the feature that the depth of the pixel on the edge of the object in the image is the same as the depth of the closer object, the corresponding disparity non-uniformity function of each type of pixel is designed respectively:

B1) Parallax inconsistency function design corresponding to the first type of pixels:

If the adjacent pixels of the current pixel are on the edge of the object in the image, the corresponding disparity non-uniformity function is:

V _{{p, q}∈N} (f _p , f _q )=λ|f _p -f _q |

In the formula: f _p represents the disparity of the current pixel p; f _q represents the disparity of the adjacent pixel q of the current pixel p; λ represents the smoothing factor, and the value is 1;

If the adjacent pixel of the current pixel is next to the edge of the object in the image, it is judged whether the adjacent pixel is in the closer object object, and if so, the corresponding parallax inconsistency function is:

V _{{p, q} ∈ N} (f _p , f _q ) = λ|f _p -f _q |;

If the adjacent pixel is in the far object object, the parallax inconsistency function is:

V _{{p, q} ∈ N} (f _p , f _q ) = ∞;

Thus, the parallax inconsistency function of the first type of pixels is designed as:

In the formula: E(·) is edge auxiliary information, E(·)=1 represents that the pixel is located on the edge of the object in the image, and E(·)=0 represents that the pixel is not on the edge of the object in the image;

B2) Parallax inconsistency function design corresponding to the second type of pixel:

If the current pixel and the adjacent pixels of the current pixel are all next to the edge of the object in the image, then the corresponding disparity non-uniformity function is:

V _{{p, q} ∈ N} (f _p , f _q ) = λ|f _p -f _q |;

If the adjacent pixel of the current pixel is on the edge of the object in the image, it is judged whether the current pixel is in the closer object, if the current pixel is in the closer object, the corresponding disparity non-uniformity function is:

V _{{p, q} ∈ N} (f _p , f _q ) = λ|f _p -f _q |;

If the current pixel is in the far object, the parallax inconsistency function is:

V _{{p, q} ∈ N} (f _p , f _q ) = ∞;

Thus, the parallax non-uniformity function of the second type of pixels is designed as:

B3) Parallax inconsistency function design corresponding to the third type of pixel:

If the current pixel and its surrounding adjacent pixels are not on the edge of the object in the image, then the parallax non-uniformity function of such pixels is:

V _{(p, q)∈N} (f _p , f _q )=λ|f _p -f _q | If E(p)=0 and E(q)=0;

C. According to the brightness non-uniformity function and the obtained parallax non-uniformity function of the three types of pixels, calculate the pixel brightness non-uniformity and parallax non-uniformity respectively, and use the energy minimization function to perform corresponding parallax estimation

D. According to the estimated disparity value, use the disparity depth conversion function to convert the disparity value into a corresponding depth value, so as to complete the depth estimation.

Compared with the prior art, the present invention has the following advantages:

Because the present invention makes full use of the depth characteristics of the object edge pixels in the image, according to whether the pixels next to the object edge in the image are in the position in the closer object object, the disparity corresponding to the divided three types of pixels is designed to be inconsistent Therefore, compared with the existing depth estimation method in the FTV system, the obtained depth map is more accurate at the edge of the object, so that the synthesized virtual view has fewer noise points at the edge of the object, effectively improving the Subjective quality and objective quality of synthetic views.

Description of drawings

Fig. 1 is a structural block diagram of functional modules in the FTV system;

Fig. 2 is the flowchart of existing depth estimation method;

Fig. 3 is a flow chart of the depth estimation method of the present invention;

FIG. 4 is a depth map obtained by using the existing depth estimation method and the depth estimation method of the present invention respectively;

Fig. 5 is a synthetic virtual view obtained by using the existing depth estimation method and the depth estimation method of the present invention respectively.

Detailed ways

Referring to Fig. 3, the depth estimation method of the present invention comprises the following steps:

Step 1, classify the pixels in the image according to the specific position of the current pixel.

1A) If the current pixel is located on the edge of the object in the image, then the current pixel is divided into the first type of pixels, i.e. the pixels on the edge;

1B) If the current pixel is next to the edge of the object in the image, and there are pixels located on the edge of the object in its surrounding adjacent pixels, then the current pixel is assigned to the second type of pixel, i.e., the pixel next to the edge;

1C) If the current pixel is on the edge of the object in the image, and its surrounding adjacent pixels are all on the edge of the object in the image, then classify the current pixel as the third type of pixel, ie non-edge pixel.

Step 2, according to the depth features of the object edge pixels in the image, the disparity inconsistency functions corresponding to the three types of pixels are designed.

The depth features of the edge pixels of the object in the image are as follows:

(a) A scene consists of a space filled with objects. Each object in the scene has its contour line, which is reflected in the image as the edge line. The edge line is the dividing line projected in the image between objects in the scene, and there are often two different object projections on both sides of the edge in the image. So the depth of the edge point is not the same as the depth of its neighbors, but not necessarily the same;

(b) In the direction of people's line of sight, objects overlap from near to far until infinity. The overlapping of objects constitutes an edge, and the edge point belongs to the closer object among the two overlapping objects, that is, the depth of the edge point is the same or similar to the depth of the closer object;

(c) Even if the edge is the boundary of the segmented object, there are still false edges in the image, that is, there are lines similar to the edge inside an object, and their depth relationship with adjacent points is the same as that of adjacent points in the same object. .

The disparity inconsistency functions corresponding to the three types of pixels are as follows:

2A) Parallax inconsistency function design of the first type of pixels

In the first type of pixels, the current pixel is on the edge of the object in the image, if the adjacent pixels of the current pixel are also on the edge of the object in the image, the parallax change of the pixels on the same edge should be slow , then the corresponding parallax inconsistency function is:

V _{(p, q}∈N} (f _p , f _q )=λ|f _p -f _q |

In the formula: f _p represents the disparity of the current pixel p; f _q represents the disparity of the adjacent pixel q of the current pixel p; λ represents the smoothing factor, and the value is 1;

If the adjacent pixels of the current pixel are beside the edge of the object in the image, it is necessary to determine whether the adjacent pixels are in the closer object. Let the current pixel be p, and its left neighbor pixel q is located next to the edge. Compare the parallax difference between the adjacent pixel q and the current pixel p, and the parallax difference between the adjacent pixel q and the adjacent pixel m to the left of the pixel q, if |f _q -f _p |≤|f _q -f _m |, then the pixel q belongs to the closer object object; otherwise, the pixel q belongs to the far object object, wherein: f _m represents the disparity of the pixel m;

If the adjacent pixel is in the closer object, since the current pixel is in the closer object, the two belong to the same object. According to the depth characteristics (b) and (c) of the edge pixels of the object in the image, the pixel p and its adjacent pixel q belong to the same depth object, and the parallax change between the two should be slow, so the corresponding parallax is inconsistent The sex function is:

V _{{p, q} ∈ N} (f _p , f _q ) = λ|f _p -f _q |;

If the pixel q is located in the far object object, according to the depth characteristics (b) and (c) of the object edge pixel in the image, it does not belong to the same depth object as the pixel p located on the object edge in the image , the disparity difference between the two is large, then the corresponding disparity non-consistency function is:

V _{{p, q} ∈ N} (f _p , f _q ) = ∞;

Thus, the parallax inconsistency function of the first type of pixels is designed as:

In the formula: E( ) is edge auxiliary information, E( )=1 represents that this pixel is located on the object edge in the image, E( )=0 then represents that this pixel is located beside the object edge in the image;

2B) Parallax inconsistency function design of the second type of pixels

In the second type of pixels, the current pixel is at the edge of the object in the image, if the adjacent pixels of the current pixel are also at the edge of the object in the image, and the parallax change between the two should be slow, then the corresponding The disparity inconsistency function of is:

V _{{p, q} ∈ N} (f _p , f _q ) = λ|f _p -f _q |;

If the adjacent pixels of the current pixel are on the edge of the object in the image, it is necessary to judge whether the current pixel is in the closer object. Let the current pixel be p, and its left adjacent pixel q is located on the edge of the object in the image; compare the parallax difference between the adjacent pixel q and the current pixel p, and the adjacent pixel q and the left side of the pixel q The parallax difference value of the adjacent pixel m, if |f _q -f _p |≤|f _q -f _m |, then the pixel p belongs to the closer object; otherwise, the pixel p belongs to the far object;

If the current pixel is in the closer object, since the adjacent pixel is in the closer object, the two belong to the same object. Also according to the depth characteristics (b) and (c) of the object edge pixels in the image, the pixel p and its adjacent pixel q belong to the same depth object, and the parallax change between the two should be slow, then the corresponding parallax The non-uniform function is:

V _{{p, q} ∈ N} (f _p , f _q ) = λ|f _p -f _q |;

If the pixel p is located in the far object object, according to the depth characteristics (b) and (c) of the object edge pixel in the image, it does not belong to the same depth object as the pixel q located on the object edge in the image , the disparity difference between the two is large, then the corresponding disparity non-consistency function is:

V _{{p, q} ∈ N} (f _p , f _q ) = ∞;

Thus, the parallax non-uniformity function of the second type of pixels is designed as:

2C) Parallax inconsistency function design of the third type of pixels

In the third type of pixels, the current pixel and its adjacent pixels are located near the edge of the object in the image, so the parallax change between the two should be slow, and the designed parallax of such pixels is inconsistent The sex function is:

V _{(p, q)∈N} (f _p , f _q )=λ|f _p −f _q |if E(p)=0 and E(q)=0.

Step 3, use the energy minimization function to estimate the disparity.

3A) Calculation of pixel brightness inconsistency

Pixel brightness inconsistency measures the brightness inconsistency between the pixel p in the current view and the pixel in the reference view (p+f _p ) under the condition of disparity f _p , and is calculated by the following formula:

${\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}<span\; class="notranslate">\; |</span>$

表示参考视图中的象素(p+f_p)的亮度值；In the formula: I _p represents the brightness value of the pixel p in the current view;

Indicates the brightness value of the pixel (p+f _p ) in the reference view;

3B) Pixel parallax inconsistency calculation

Pixel disparity inconsistency measures the inconsistency of the disparity between the disparity of the pixel p in the current view when the disparity of the pixel p is f _p , and the disparity of its adjacent pixel q is f _q . The inconsistency function V _{{p, q}∈N} (f _p , f _q ) of the prime is calculated;

3C) Disparity Estimation

The disparity is estimated according to the energy minimization function by using the computed pixel brightness inconsistency and pixel disparity inconsistency.

The energy minimization function defines the best disparity combination f that can minimize the energy function E, as shown in the following formula:

E(f)＝E _data (f)+E _smooth (f)

其中P表示当前视图中的象素集合；

其中N表示邻近象素对的集合。In the formula:

Where P represents the set of pixels in the current view;

where N represents the set of adjacent pixel pairs.

Step 4, using the disparity value to calculate the depth value.

According to the disparity value estimated in step 3, using the disparity depth conversion function, the disparity value is converted into a corresponding depth value to complete the depth estimation;

The parallax depth conversion function is as follows:

In the formula: d represents the estimated parallax value; Z represents the depth value; I represents the distance between the cameras; f is the focal length of the camera lens; Z _near and Z _far represent the depth value and The depth value of the farthest plane.

Effect of the present invention can further prove by following experiment:

1. Experimental conditions

The present invention adopts the depth estimation reference software and the virtual view synthesis reference software provided by the moving picture expert group MPEG organization as the experimental platform, and integrates the depth estimation method based on edge pixel features of the present invention with the existing depth estimation method in the FTV system respectively into the depth estimation reference software. At the same time, three groups of multi-viewpoint video sequences provided by MPEG organization are used in the experiment, namely champagne_tower, book_arrival and newspaper. In order to verify the effectiveness of the method of the present invention, the quality of the depth map obtained by the two methods is compared, and the subjective and objective quality of the virtual view synthesized at the receiving end of the FTV system is compared and analyzed.

2. Experimental content

First, the existing depth estimation method and the depth estimation method based on edge pixel features in the present invention are used to estimate the depth data respectively, and the obtained depth map of the champagne_tower sequence is shown in FIG. 4 . Fig. 4(a) is the depth map result of the existing depth estimation method, and Fig. 4(b) is the depth map result of the depth estimation method of the present invention.

Secondly, the virtual view is synthesized by using the collected multi-viewpoint view and the depth map of the corresponding view estimated by the two methods, and the obtained synthetic virtual view of the champagne_tower sequence is shown in Figure 5. Figure 5(a) is the result of virtual view synthesized by using the existing depth estimation method, and Figure 5(b) is the result of virtual view synthesized by using the depth estimation method of the present invention.

Finally, calculate the peak signal-to-noise ratio (PSNR) of the virtual view obtained by using the two depth estimation methods, and the results are shown in Table 1.

Table 1 PSNR results of synthetic virtual view based on two methods

3. Experimental analysis

As can be seen from FIG. 4 , since the depth characteristics of the object edge pixels in the image are fully utilized, the depth map obtained by the depth estimation method of the present invention is more accurate at the edge of the object, which greatly reduces image noise.

It can be seen from FIG. 5 that since the depth estimation method of the present invention obtains a more accurate depth map, the subjective quality of the synthetic virtual view edge is greatly improved.

It can be seen from Table 1 that compared with the existing depth estimation method, the objective quality of the virtual view obtained by the method of the present invention has been improved to varying degrees, and the average can reach 0.32dB.

Claims (3)

1. the depth estimation method based on edge pixel features comprises the steps:

A. the pixel in the image is classified, the pixel in the image is divided three classes: the first kind is a pixel on the edge, and second type is the other pixel in edge, and the 3rd type is non-edge pixel;

B. according to the degree of depth characteristic identical that is in the pixel on the object edge in the image, design each type pixel parallax non-uniformity function corresponding respectively with the degree of depth of closer objects:

B1) the corresponding parallax non-uniformity function design of first kind pixel:

If the adjacent image point of present picture element is on the object edge in the image, then parallax non-uniformity function corresponding is:

V _{p，q}∈N(f _p，f _q)＝λ|f _p-f _q|

In the formula: f _pThe parallax of expression present picture element p; f _qThe parallax of the adjacent image point q of expression present picture element p; λ representes smoothing factor, and value is 1;

If by the object edge of adjacent image point in image of present picture element, judge whether adjacent image point is in the closer objects object, if then parallax non-uniformity function corresponding is:

V _{p，q}∈N(f _p，f _q)＝λ|f _p-f _q|；

If adjacent image point is in than in the far object object, then parallax non-uniformity function is:

V _{p，q}∈N(f _p，f _q)＝∞；

Thereby the parallax non-uniformity function of designing first kind pixel is:

In the formula: E () is the edge supplementary, and E ()=1 this pixel of expression is arranged on the object edge of image, and E ()=0 this pixel of expression is in by the object edge in the image;

B2) second type of corresponding parallax non-uniformity function design of pixel:

If all in image by the object edge, then parallax non-uniformity function corresponding is the adjacent image point of present picture element and present picture element:

V _{p，q}∈N(f _p，f _q)＝λ|f _p-f _q|；

If the adjacent image point of present picture element on the object edge, judges whether present picture element is in the closer objects object in image, if present picture element is in the closer objects object, parallax non-uniformity function corresponding is:

V _(p，q}∈N(f _p，f _q)＝λ|f _p-f _q|；

If present picture element is in than in the far object object, then parallax non-uniformity function is:

V _{p，q}∈N(f _p，f _q)＝∞；

Thereby the parallax non-uniformity function of designing second type of pixel is:

B3) the 3rd type of corresponding parallax non-uniformity function design of pixel:

On every side adjacent image point is not all in image on the object edge if present picture element is with it, and then the parallax non-uniformity function of this type of pixel is:

V _{(p, q) ∈ N}(f _p, f _q)=λ | f _p-f _q| if E (p)=0 and E (q)=0;

C. according to the parallax non-uniformity function of brightness nonuniformity function and resulting three types of pixels, calculating pixel brightness nonuniformity and parallax non-uniformity utilize the energy minimization function to carry out corresponding disparity estimation respectively,

Said brightness nonuniformity function is:

In the formula: I _pThe brightness value of the pixel p of ' expression in front view;

Pixel (p+f in the expression reference-view _p) brightness value;

D. according to the parallax value of estimating to obtain, utilize parallax degree of depth transfer function, convert parallax value into corresponding depth value, with the completion estimation of Depth,

Described parallax degree of depth transfer function Z is shown below:

In the formula: d representes to estimate the parallax value that obtains; Z representes depth value; I representes the video camera spacing; The focal length of f camera lens; Z _NearWith Z _FarDepth value and the depth value on maximum distance plane of representing the minimum distance plane of the object distance video camera in the three-dimensional scenic respectively.

2. the depth estimation method based on edge pixel features according to claim 1; Wherein steps A is described classifies to the pixel in the image; Be to confirm,, then this pixel belonged to first kind pixel if present picture element is in the image on the object edge according to the position of present picture element; If present picture element is not in image on the object edge, and there is the pixel on the object edge in the image in pixel around it, then this pixel is belonged to second type of pixel; If present picture element is not in image on the object edge, and pixel all on the object edge, then is not classified as the 3rd type of pixel with this pixel around it in image.

3. the depth estimation method based on edge pixel features according to claim 1, wherein saidly judge that whether present picture element is in the closer objects object, carries out as follows:

Make present picture element p be positioned on the edge, its left adjacent image point q is positioned at by the edge;

The disparity difference of the disparity difference of comparison adjacent image point q and present picture element p and the adjacent image point m of adjacent image point q and pixel q left, if | f _q-f _p|≤| f _q-f _m|, then pixel q belongs in the closer objects object, otherwise pixel q belongs to than in the far object object, wherein: f _mThe parallax of expression pixel m.

Images