CN106951829A - A kind of notable method for checking object of video based on minimum spanning tree - Google Patents

Abstract

The invention discloses a kind of notable method for checking object of the video based on minimum spanning tree, including step：Super-pixel segmentation, the distance of each super-pixel of calculating to border, and the minimum spanning tree of structural map picture are carried out to input picture；The outline information extracted by the image minimum spanning tree range conversion rule of foundation and using quick profile testing method is combined, and is extracted and is obtained preliminary notable figure；Fuzzy Processing is carried out to the on-fixed outline of the preliminary notable figure of gained using the fuzzy aberration histogram feature of foundation, clear-cut notable figure is obtained；Notable object uses the multiple scale detecting for the sampling feature grid for being adapted to pedestrian's window in the notable figure clear-cut to gained, and is incorporated into final notable figure.The present invention can not be influenceed by the external interference factor such as illumination variation, can effectively, the human body behavior characteristic characterization of strong robustness, amount of calculation is small, real-time is high and has preferably notable object detection effect.

Description

A kind of notable method for checking object of video based on minimum spanning tree

Technical field

The present invention relates to a kind of notable method for checking object of the video based on minimum spanning tree, belong to the technology of video detection Field.

Background technology

Target detection technique is the heat subject that computer vision is studied with area of pattern recognition, is learned both at home and abroad The common concern of person, is with a wide range of applications.The target to be realized of target detection is, with computer by target object from Detected in two dimensional image containing other object backgrounds for the same target in computer vision field, significant property detection It is that the things for most attracting the mankind in the mankind is predicted by simulating human visual system so that effectively and can efficiently reduce For the complexity of the scene analysis further handled.It is redirected in image, image and video compress, Object identifying, image Classification and picture piece together wait field extensive use have studied some time.

The detection of target is the basis of video monitoring system graphical analysis, in the case where accurately detecting target, just may be used To extract validity feature.By detect and track, the attitude and behavioral parameters of moving target can be easily obtained, is image reason Solution provides technical support.Detection is also the basis of vision measurement.Detection always links together with tracking, is accounted in video monitoring There is consequence.From being born from vision technique, target detection is just paid much attention to, and have accumulated numerous studies Achievement, meanwhile, the detection of moving target is also the problem of being rich in challenge in vision research, to there is many theoretical and actual technologies Difficult point needs to solve.Although the research of target detection makes some progress, but still there is conspicuousness target in illumination difference It is different, caused missing inspection when complex background and dimensional variation, flase drop and it is inefficient the problem of, to find out its cause, mainly in research In there is viewpoint change, the difficulty of several aspects such as complex background and light differential or challenge.

The content of the invention

The technical problems to be solved by the invention are to overcome the deficiencies in the prior art to be based on minimum spanning tree there is provided one kind The notable method for checking object of video, solve target between because conspicuousness target is in light differential, complex background and dimensional variation When caused missing inspection, flase drop and it is inefficient the problem of.

It is of the invention specific using following technical scheme solution above-mentioned technical problem：

A kind of notable method for checking object of video based on minimum spanning tree, comprises the following steps：

Step A, super-pixel segmentation is carried out to input picture, calculate each super-pixel to the distance on border, and structural map picture Minimum spanning tree；Outside by the image minimum spanning tree range conversion rule of foundation and using the extraction of quick profile testing method Profile information is combined, and is extracted and is obtained preliminary notable figure；

Step B, using the fuzzy aberration histogram feature of foundation to the on-fixed outline of preliminary notable figure obtained by step A Fuzzy Processing is carried out, clear-cut notable figure is obtained；

Significantly object is special using the sampling for being adapted to pedestrian's window in step C, the notable figure clear-cut to step B gained The multiple scale detecting of grid is levied, and is incorporated into final notable figure.

Further, as a preferred technical solution of the present invention, set up in the step A image minimum spanning tree away from From transformation rule using bottom-up traversal and top-down traversal.

Further, fuzzy aberration histogram is set up as a preferred technical solution of the present invention, in the step A special Levy, be specially:Image color intensity I is quantified as W grades, is fallen into a trap in local adjacent area and calculates color difference；

The color difference lease making is crossed into Gauss member function and carries out obfuscation, and color difference Nogata is calculated in regional area Figure；

Fuzzy color difference histogram feature is formed using fuzzy c-means clusters and color histogram of difference.

Further, as a preferred technical solution of the present invention, quick profile testing method is utilized in the step B The outline information of extraction, be specially:

Input picture is resolved into some components；

Region of the morphological operator generation with profile information is performed to each component；

The profile in each region is calculated, and chooses the conspicuousness that wherein each pixel is calculated after maximum profile.

Further, as a preferred technical solution of the present invention, the profile in each region is calculated using formula：

Wherein, r₀It is region r outline, | r₀| it is r₀The length of profile, and E (x) is pixel x profile confidence.

The present invention uses above-mentioned technical proposal, can produce following technique effect：

The method that the present invention is provided, carries out super-pixel segmentation to image first, then calculates each super-pixel and arrive border Distance, and minimum tree is generated by range conversion, notable object is tentatively extracted, is then extracted by quick profile testing method Outline information, finally, Fuzzy Processing generation notable figure is carried out with fuzzy aberration histogram feature and is incited somebody to action to on-fixed outline It is attached to final notable figure.Real-time notable method for checking object of the invention based on minimum spanning tree can not be by illumination variation Etc. the influence of external interference factor, a kind of effective, human body behavior characteristic characterization of strong robustness is searched out, passes through one kind calculating Measure small, real-time high and there is the algorithm of preferable forecast function mainly to solve between target because conspicuousness target is in light differential, it is multiple Caused missing inspection when miscellaneous background and dimensional variation, flase drop and it is inefficient the problem of.The detection speed of the present invention is compared to existing Method is higher, and the performance of comprehensive detection speed and the accuracy rate present invention are better than other existing methods.

Brief description of the drawings

Fig. 1 is the schematic flow sheet of the video notable method for checking object of the invention based on minimum spanning tree.

Fig. 2 (a) for image superpixel and edge in the image minimum spanning tree construction of the present invention by adjacent pixel it Between color/intensity difference come the plan that weights；Fig. 2 (b) is to be shown by sequentially removing the side with big weight to construct MST It is intended to；Fig. 2 (c) is that the distance of two hollow nodes defines schematic diagram on tree path.

Fig. 3 (a) is starting stage minimum spanning tree schematic diagram in distance transform algorithm of the present invention；Fig. 3 (b) is in the present invention The schematic diagram of traversal is updated from bottom to top；Fig. 3 (c) is the top-down schematic diagram for updating traversal in the present invention.

Fig. 4 is the schematic diagram of the conspicuousness detection process proposed by the present invention based on minimum spanning tree.

Fig. 5 is adapted to the schematic diagram of the multiple scale detecting of the sampling feature grid of pedestrian's window for the present invention.

Embodiment

Embodiments of the present invention are described with reference to Figure of description.

As shown in figure 1, the present invention proposes a kind of notable method for checking object of video based on minimum spanning tree, this method tool Body comprises the following steps：

Step A, super-pixel segmentation is carried out to input picture, calculate each super-pixel to the distance on border, and structural map picture Minimum spanning tree；Outside by the image minimum spanning tree range conversion rule of foundation and using the extraction of quick profile testing method Profile information is combined, and is extracted and is obtained preliminary notable figure.

Specifically, the present invention is defined first carries out minimum spanning tree construction to input picture, and the present invention regard image I as mark The undirected plans of the connection of standard 4, its interior joint be all edge between image superpixel, and neighbouring super pixels by color/ Intensity difference is weighted, and it is absolute gradient, can according to Kruskal algorithm by sequentially remove side with big weight come Minimum spanning tree is constructed, the remaining edge connected by all pixels is left and is used as tree.More specifically, it is adjacent a pair to make p and q Super-pixel, the weight between p and q is：

ω (p, q)=ω (q, p)=| I (p)-I (q) |¹ (1)

The present embodiment shows the simple examples of 5 × 5 pixels in Fig. 2 (a).Fig. 2 (a) is that its interior joint is image superpixel And the plan that edge is weighted by the color/intensity difference between adjacent pixel.Fig. 2 (b) is by sequentially removing tool There is the side of big weight to construct minimum spanning tree.Fig. 2 (c) is that the distance of two red nodes is defined on tree path, i.e. dotted line Represent.During construction is set, if two neighbouring super pixels have big gradient, the edge between the two nodes may quilt Remove, and it will be long to the path of another node to travel through a node, such as shown in Fig. 2 (c).As a result, for the two Distance on node, minimum spanning tree will be big.On the contrary, if the outward appearance of two nodes is similar, they may be in minimum Connected in spanning tree, and apart from will be very short.

The present invention carries out range conversion using minimum spanning tree, is that a kind of minimum spanning tree based on a width picture represents Novel range conversion.The algorithm proposed can apply two kinds of distance metrics of geodesic distance and potential barrier distance.The present invention is proposed Shortest path is found on minimum spanning tree.The range conversion based on minimum spanning tree proposed is the distance defined in the tree Exact method.It includes two traversals：Bottom-up traversal and top-down traversal.Fig. 3 (a) gives to figure (c) and used The schematic diagram of geodesic distance.The use of potential barrier distance is similar.Give one group of seed node S, the present invention by seed node away from 0 is initialized as from value, the distance value of every other node ∞ is initialized as, shown in such as Fig. 3 (a).Fig. 3 (b) and (c) are respectively Traversal and top-down traversal from bottom to top.

For time from bottom to top, the present invention updates with following formula the distance value of its father node since leaf node：

M (p)=min { M (p), f (ζ_υ∪p)} (2)

Wherein, p represents υ father node, ζ_υυ is connected to the current best path of its nearest seed node, ζ by expression_υ∪ P represents identical path plus a step further to his father p.Due to nearest seed node p can come from its bottom or from its Top, from bottom, test possible solution formula 2.If p has multiple child nodes, formula (2) will be directed to each child node It is estimated, and stores minimum range.Fig. 3 (b) is an example of root node.In brief, in each node with wide Degree first search (BFS) is sequentially updated processing.Bottom-up traversal reaches root node until it.

Since root node, top-down process is similar.For each node, the present invention accesses its child node simultaneously Their distance value is updated with following formula：

M (υ)=min { M (υ), f (ζ_p∪υ)} (3)

Wherein, not only from top, test possible solution formula (3), and can propagate potential from other branches More preferable solution.It can be seen that the diagram in Fig. 3 (b).After process from bottom to top, many nodes and seed node Distance be still ∞.Nearest seed node to a certain node is likely located at top or in node separation from the upper side Bifurcation.Bottom-up after, it is expected that spliting node records the optimum distance value from one of its branch.From upper and Under process in, optimal solution will propagate down into other branches, shown in such as Fig. 3 (c).

The distance between super-pixel computational methods can be determined by traveling through mode according to above two, can be obtained according to this method To the distance between two super-pixel nodes value.Specially：First, node υ₁Distance by seed node υ₂Uniquely determine.This Be based on the fact that：Seed node is advanced through to have an opportunity the distance for increasing geodesic distance and potential barrier distance, therefore υ₂It is υ₁'s Optimal seed node.For geodesic distance, row is further further added by an absolute gradient and calculated to its distance, so update Distance is not reduce.For potential barrier distance, the maximum and minimum value of each node are tracked.One node of traversal can update most Big value or minimum value, so that increase barrier value, therefore final distance is not also reduced.This means cross over seed node or more novel species The distance of child node will not provide more preferable result, therefore be unnecessary.

On this basis, the present invention can again be carried out according to the distance between the above-mentioned two super-pixel nodes calculated value The process of range conversion rule.Setting up image minimum spanning tree range conversion procedure of rule is specially：

(1) bottom-up traversal is performed, top-down traversal is then performed, optimum distance conversion is obtained.

(2) if seed node is unique seed node and the node on the root node of subtree, subtree Range conversion is uniquely determined by the seed node.Will from top to bottom during obtain corresponding range conversion.

(3) during traveling through, renewal step of the seed node in formula (2) and formula (3) can be ignored.

Measured Boundary of the present invention is connective, in order to supplement the shortage of range conversion, introduces a border auxiliary figure, using fast The outline information that fast profile testing method is extracted, by color similarity survey calculation node by node, to improve conspicuousness detection Quality.Its detailed process is as follows：

Input picture is resolved into some components, morphological operator generation is performed to each component has similar appearance and wheel The region that wide information is preserved.If saliency object region is surrounded completely by other regions, the encirclement degree in region is 1, And if located in then the encirclement in region is 0 in image boundary, this marking area being pointed in image boundary is unfair.It is based on Encircled area tends to the observation with strong outline, and the present invention proposes appropriateness measurement to determine the encirclement degree in region, and Measured by the mean profile confidence level of its outline.Profile confidence map E is produced by quick profile testing method.For cloth Each join domain r in your Mapping B, profile is calculated by below equation：

Wherein, r₀It is region r outline, | r₀| it is r₀The length of profile, and E (x) is pixel x profile confidence.

In order to further suppress the encirclement in the region being located in image boundary, peer-to-peer (4) addition punishment：

Wherein, I_borderIt is provided in the pixel on framing mask.Outline pixel is more on image boundary region, punishes Penalize Xiang Yue great.If the region is away from image boundary, penalty term is zero.Finally, the present invention calculates each by below equation Pixel x conspicuousness：

S_S(x)=max { s₁(x),...,s_|B′|(x)} (6)

Step B, using the fuzzy aberration histogram feature of foundation to the on-fixed outline of preliminary notable figure obtained by step A Fuzzy Processing is carried out, clear-cut notable figure is obtained, as shown in Figure 4.

Image color intensity I (u, v, ch) is first quantified as W grades, I ∈ { 0,1 ..., W-1 }, (u, v) is coordinate, and ch is color Multimedia message road.The intensity of the ch color channel is centered on (u, v), and I (p, q, ch) represents the intensity of adjacent pixel.

First, the color histogram of image is calculated：Hg (i, j, k) be pixel using centered on coordinate (u, v) size as M × The possibility of M regional area, be specially：

Secondly, color histogram of difference is calculated：The homogeneous color difference of two pixels can be perceived, in a small part Color difference point is calculated in adjacent area R × R：

Wherein, color difference d carries out obfuscation with Gauss member function：

Then, color histogram of difference is calculated in regional area M × M：

Finally, the fuzzy aberration histogram of image is obtained：

Fuzzy aberration histogram feature is formed using fuzzy c-means clusters and color histogram of difference.

Fuzzy c-means is clustered N number of local histogram X={ x₁,x₂,...,x_NBe divided into fuzzy member function c with υ_iCentered on cluster, can with iteration minimize cost function obtain：

The FCDH vectors h of c dimensions：

h_c×1=γ_c×NH_N×1 (12)

Therefore, Fuzzy Processing is carried out to the on-fixed outline of preliminary notable figure using fuzzy aberration histogram feature to obtain Clear-cut notable figure, wherein clear-cut notable figure has multiple different size of notable objects.

Significantly object is special using the sampling for being adapted to pedestrian's window in step C, the notable figure clear-cut to step B gained The multiple scale detecting of grid is levied, and is obtained final notable figure.

Conventional method uses single grader for the pedestrian dummy of fixed dimension, and dependent on multiple yardsticks Characteristics of image is recalculated or approximately.In order to detect the pedestrian at multiple yardsticks, the filtering that the present invention calculates original scale is led to Road, and using 1.07 scale factor in multiple scale dimension applications sliding windows, every about 10 yardsticks of half of octave utilize many chis Degree is applied in the video detection of the present invention, improve conspicuousness target in video far and near change when robustness.Such as Fig. 5 institutes Show, the present invention by extracting characteristic of division from the channel sample of filtering in a grid-like fashion.Grid adaptation is in the big of detection window It is small.So, for the pedestrian of any size, the feature of identical quantity is obtained using different grid intervals.In classifier training In the case of, only calculate characteristics of image, and uncomfortable whole pedestrian image in original scale.Feature is sampled in fig. 5 it is shown that wherein Multiple different size of notable object detections can be gone out in the clear-cut notable figure of gained, and be attached in notable figure most Display eventually.

For sliding window, the present invention uses following adaptive step rate：1/16 and vertical window of horizontal window width The 1/16 of height.By the way that window center regulation is arrived between row 140 and 300, search space reduces 35%.

The channel characteristics of filtering are not Scale invariants, and different scale pedestrian by with different expressions.Pedestrian For different illuminations, direction or circumstance of occlusion also have a significant different expression, and be made up of thousands of weak learners Single enhancing grader remains able to provide consistent result.The present invention being capable of independent study for different size of pedestrian Related classification feature represents for different pedestrians.

To sum up, the real-time notable method for checking object of the invention based on minimum spanning tree, can not be outer by illumination variation etc. The influence of boundary's disturbing factor, searches out a kind of effective, human body behavior characteristic characterization of strong robustness.The detection speed of the present invention Higher compared to existing method, the performance of comprehensive detection speed and the accuracy rate present invention are better than other existing methods.

Embodiments of the present invention are explained in detail above in conjunction with accompanying drawing, but the present invention is not limited to above-mentioned implementation Mode, can also be on the premise of present inventive concept not be departed from the knowledge that those of ordinary skill in the art possess Make a variety of changes.

Claims (5)

1. the notable method for checking object of a kind of video based on minimum spanning tree, it is characterised in that comprise the following steps：

Step A, super-pixel segmentation is carried out to input picture, calculate each super-pixel to the distance on border, and structural map picture is most Small spanning tree；The outline extracted by the image minimum spanning tree range conversion rule of foundation and using quick profile testing method Information is combined, and is extracted and is obtained preliminary notable figure；

Step B, fuzzy aberration histogram feature is utilized to carry out fuzzy place to the on-fixed outline of preliminary notable figure obtained by step A Reason, obtains clear-cut notable figure；

Significantly object is using the sampling feature net for being adapted to pedestrian's window in step C, the notable figure clear-cut to step B gained The multiple scale detecting of lattice, obtains final notable figure.

2. the notable method for checking object of video according to claim 1 based on minimum spanning tree, it is characterised in that the step Image minimum spanning tree range conversion rule is set up in rapid A using bottom-up traversal and top-down traversal.

3. the notable method for checking object of video according to claim 1 based on minimum spanning tree, it is characterised in that：The step Outline information is extracted using quick profile testing method in rapid A, is specially：

Input picture is resolved into some components；

Region of the morphological operator generation with profile information is performed to each component；

The profile in each region is calculated, and chooses the conspicuousness that wherein each pixel is calculated after maximum profile.

4. the notable method for checking object of video according to claim 1 based on minimum spanning tree, it is characterised in that the step Fuzzy aberration histogram feature is set up in rapid B, is specially：

Image color intensity I is quantified as W grades, is fallen into a trap in local adjacent area and calculates color difference；

The color difference lease making is crossed into Gauss member function and carries out obfuscation, and color histogram of difference is calculated in regional area；

Fuzzy color difference histogram feature is formed using fuzzy c-means clusters and color histogram of difference.

In the histogrammic method of the fuzzy color difference of outline application of well-marked target, clear-cut conspicuousness target is obtained.

5. the notable method for checking object of video according to claim 3 based on minimum spanning tree, it is characterised in that：It is described every The profile in individual region is calculated using formula：

$S_i\left(r\right)=\frac{{\Σ}_{x\∈r_0}E\left(x\right)}{\left|r_0\right|}$

Wherein, r₀It is region r outline, | r₀| it is r₀The length of profile, and E (x) is pixel x profile confidence.