CN105405138B

CN105405138B - Waterborne target tracking based on conspicuousness detection

Info

Publication number: CN105405138B
Application number: CN201510761806.7A
Authority: CN
Inventors: 王贺升; 陈卫东; 智绪浩
Original assignee: Shanghai Jiaotong University
Current assignee: Shanghai Jiaotong University
Priority date: 2015-11-10
Filing date: 2015-11-10
Publication date: 2018-05-29
Anticipated expiration: 2035-11-10
Also published as: CN105405138A

Abstract

The present invention provides a kind of waterborne target trackings based on conspicuousness detection, feature is to detect to provide the ROI of target following and location verification information first with waterfront line, afterwards using incremental encoding length as the index for weighing image energy information, represent the rare degree of feature in image, the target in notable feature figure is extracted using water surface feature and target signature, so as to complete waterborne target track algorithm.It is compared with common target tracking algorism, the present invention is more applicable for waterborne target tracking, has the characteristics of real-time is outstanding, illumination adaptive.

Description

Water surface target tracking method based on significance detection

Technical Field

The invention relates to a water surface target tracking method based on significance detection, in particular to a video target tracking method applied to water image capture and video data analysis.

Background

The video tracking algorithm takes an image sequence as input, and takes attributes such as size and position of a target in an image as output. In an ideal situation, the output target attribute is accurate and real-time, however, in the real world, due to the fact that the background changes and various interferences exist, the ideal state is often difficult to achieve, and therefore finding a real-time algorithm which is suitable for the current environment and has certain self-adaptability to illumination and background changes is the key of target tracking.

The following algorithms are commonly used at present:

the Meanshift algorithm is a method for matching images of a fixed-size area by using color information, can calculate a probability density function value according to a sample point without relying on prior knowledge, calculate the similarity of a target area and a candidate area, calculate a Meanshift vector, and continuously perform movement and iterative calculation on the candidate area until a target is found or the iterative times limit is reached.

Background subtraction method: the method is a method for identifying a moving object by utilizing the difference between a current image and a background image, and the acquisition and the update of a background model are key technologies of the method. A simple way to acquire background images is to capture background images when no objects are present in the scene, and such artificial, non-adaptive methods acquire background images that are only suitable for short-time video surveillance. Most algorithms have abandoned this non-adaptive background image estimation method. When the scene environment is not very complicated, the estimation of the background image in the scene can be finished by using statistical filtering, and a correct background estimation image can be obtained under most conditions. The background estimation method based on the Gaussian statistical model can accurately estimate the background model in a scene with part of constantly changing regions, but the calculation is complex and the real-time requirement cannot be met.

The classification-based tracking method is to convert the target tracking problem into a continuous target detection problem, and the target detection is realized by classifying different image areas into targets or backgrounds through a classifier.

The visual perception on the water surface has special technical difficulties, the influence of illumination on the water surface is large, the directions of sunlight illumination are different at different times of the same day, the reflecting intensity and the directions on the water surface are different, the influence of different weather on the light intensity on the water surface is also large, and the visual information processing algorithm is required to have high adaptability. In the process of target detection, when the target distance is far, the degree of distinction from the background is not large enough, so that image enhancement and filtering are required to extract target information.

Disclosure of Invention

Aiming at the defects and the water surface characteristics in the prior art, the invention aims to provide a water surface target tracking method based on significance detection, which has self-adaptability to illumination and can meet the real-time requirement.

The method firstly utilizes the water shoreline detection to obtain the specific position of the water shoreline, thereby providing physical environment information and further providing position verification information and ROI (region of interest) for the detection of the target. And carrying out compressed sensing on the image by using a sparse dictionary, and then measuring the rarity of the image by using the incremental coding length so as to obtain the characteristics of the saliency map.

The invention provides a water surface target tracking method based on significance detection, which comprises a water bank line detection step;

the water shoreline detection step comprises the following steps:

step A1: down-sampling the image to obtain a down-sampled image;

step A2: converting the RGB space of the image obtained by down-sampling into HSV space; setting a hue threshold value T _ h, a saturation threshold value T _ s and a brightness threshold value T _ v;

determining the area of which the hue component h is smaller than a hue threshold value T _ h or the saturation component s is smaller than a saturation threshold value T _ s in the image as a water surface area or a sky area to be distinguished;

if the brightness component v of the water surface or the sky area to be distinguished is smaller than a brightness threshold value T _ v, judging the water surface or the sky area to be distinguished as a water shoreline area, otherwise, judging the water surface or the sky area to be distinguished as a sky area;

step A3: carrying out edge detection on the water bank line region image;

step A4: and carrying out Hough transform on the detected edge to obtain a water shoreline.

Preferably, a significance detection step is further included;

the significance detection step comprises the following steps:

step B1: obtaining an ROI (region of interest) detected in the salient region according to the water shoreline;

and step B2: obtaining the incremental coding length, wherein the calculation formula is as follows:

ICL(p _i )＝-H(p)-p _i -logp _i -p _i logp _i

wherein, ICL (p) _i ) Denotes with respect to p _i The delta code length of (c); p is a radical of _i The ith row of the probability density function p is represented, i having a value range of [1, n ]](ii) a The probability density function p is defined as: p = | p ₁ ,p ₂ ,…,p _n | ^T ；

Wherein H (p) is information entropy, and n represents the row number of the probability density function p;

wherein k is a positive integer and has a value range of [1,192]；w _i Belongs to W, the value range of i is [0,192](ii) a W is a filter function, W = A ^-1 A is the sparse basis of the ROI of interest, W = [ W = _1, w _2, …,w ₁₉₂ ] ^T ；

x ^k E.x, X denotes the sampling matrix, X = [ X ] ¹ ,x ² ,…,x ^k ,…](ii) a The sampling matrix X is an image matrix of the region of interest ROI subjected to vectorization;

and step B3: obtaining a vector diagram M of the salient region, wherein the calculation formula is as follows:

M＝[m ₁ ,m ₂ ,…,m _n ]

wherein m is _k The k column of the vector diagram M representing the salient region, k having a value range of [1, n]N is the number of columns of the vector map M of the salient region; s represents a salient feature, the complete set of S is {1,2, \8230;, n }, S = { i | ICL (p) _i )>0}；

Wherein p is _j The j-th column of the probability density function p is represented, and the value range of j is [1, n ]]；ICL(p _j ) Denotes with respect to p _j The delta code length of (c);

and step B4: and converting the vector image M of the salient region into a two-dimensional image matrix M' as a salient matrix detection image.

Preferably, the method further comprises a target extraction step;

the target extraction step comprises the following steps:

step C1: for the current frame, assuming that the target position of the previous frame is Pn-1, the current salient matrix detection graph M' totally has k high-brightness areas, marking the ith connected high-brightness area as Wi, obtaining the position coordinate Pi of the center point of Wi, and obtaining the maximum gray value PIX of the pixels in Wi _i And possesses a maximum gray value PIX _i The position information Mi of the pixel of (1); wherein, i =1,2, \8230;, k;

and step C2: filtering each communicated highlight area Wi to filter out invalid highlight areas; the following filtration conditions were set:

-defining a minimum number of pixels N _min If Wi has pixel number N _i Less than N _min Filtering the Wi;

-defining a minimum grey value PIXmin if the maximum grey value PIX of Wi is _i Less than PIX _min Then Wi is filtered;

-polygon approximation of Wi, filtering Wi if the approximation is a concave polygon;

and C3: filtering out invalid highlight areas through the step C2, and then filtering out the position coordinates Pi of the center point of each Wi and the distance D of the target position of the previous frame _i And (3) calculating:

wherein x is _n-1 Abscissa, y, representing the position of the target in the previous frame _n-1 Ordinate, x, representing the position of the object in the previous frame _i Abscissa, y, representing the position of the center point of Wi _i A vertical coordinate representing a position of a center point of Wi;

selecting each D _i Minimum distance Dmin of (1);

if the minimum distance Dmin is smaller than the set maximum distance threshold DT, the connected highlight area Wi currently having the minimum distance Dmin is determined as a target area;

and if the minimum distance Dmin is greater than or equal to the set maximum distance threshold DT, the current frame target is considered to be lost, and the current target position adopts the previous frame target position Pn-1.

Preferably, the hue threshold T _ h =50, the saturation threshold T _ s =30, and the brightness threshold T _ v =180.

Compared with the prior art, the invention has the following beneficial effects:

the method is applied to the tracking of the water surface target by combining the water shoreline detection and the saliency detection under the condition of fully considering the characteristics of the water surface image. Firstly, preprocessing and color space conversion are carried out on an original image, the influence of illumination is reduced, a water shoreline can be extracted by utilizing an edge detection algorithm and a straight line detection algorithm, then an ROI is determined by utilizing the position of the water shoreline, and the next water surface target tracking work is continued. The water surface target tracking is to use the increment coding length to represent the salient information, and invalid information filtering and target extraction are carried out in the salient image, so that the target tracking process is completed.

Drawings

Other features, objects and advantages of the invention will become more apparent upon reading of the detailed description of non-limiting embodiments with reference to the following drawings:

FIG. 1 is a Gaussian pyramid;

FIG. 2 is a view of a ring-shaped suppression area, where r ₁ Denotes the inner circle radius, r ₂ Represents the outer circle radius;

FIG. 3 is a water bank line detection process;

FIG. 4 is a traversal operator;

FIG. 5 is a saliency calculation flow;

fig. 6 is a target information extraction flow.

Detailed Description

The present invention will be described in detail with reference to specific examples. The following examples will aid those skilled in the art in further understanding the present invention, but are not intended to limit the invention in any manner. It should be noted that it would be obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit of the invention. All falling within the scope of the invention.

The invention discloses a water surface target tracking algorithm based on significance detection, which is characterized in that firstly, the water bank line detection is utilized to provide ROI and position verification information of target tracking, then, the increment coding length is used as an index for measuring image energy information to represent the rarity degree of features in an image, and the water surface features and the target features are utilized to extract targets in a significant feature map, so that the water surface target tracking algorithm is completed. Compared with a common target tracking algorithm, the method is more suitable for tracking the water surface target and has the characteristics of excellent real-time performance and illumination adaptivity.

the water shoreline detection step comprises the following steps:

step A1: down-sampling the image to obtain a down-sampled image;

step A3: carrying out edge detection on the water bank line region image;

The water surface target tracking method based on the significance detection further comprises a significance detection step;

the significance detection step comprises the following steps:

step B1: obtaining an ROI (region of interest) detected in the salient region according to the water bank line; since the aquatic target does not appear on the shore, the water surface area can be obtained according to the water shoreline as the ROI for saliency detection.

ICL(p _i )＝-H(p)-p _i -logp _i -p _i logp _i

wherein, ICL (p) _i ) Denotes with respect to p _i The delta code length of (d); p is a radical of _i The ith column, i, representing the probability density function p, has a value in the range of [1, n](ii) a The probability density function p is defined as: p = | p ₁ ,p ₂ ,…,p _n | ^T ；

H (p) is information entropy, and n represents the column number of a probability density function p;

wherein k is a positive integer and has a value range of [1,192]；w _i Belongs to W, the value range of i is [0,192](ii) a W is a filter function, W = A ^-1 A is the sparse basis of the ROI of interest, W = [ W = ₁ ,w ₂ ,…,w ₁₉₂ ] ^T ；

x ^k E.x, X denotes the sampling matrix, X = [ X = ¹ ,x ² ,…,x ^k ,…](ii) a The sampling matrix X is an image matrix of the region of interest ROI subjected to vectorization;

M＝[m ₁ ,m ₂ ,…,m _n ]

wherein m is _k The k column of the vector diagram M representing the salient region, k having a value range of [1, n]N is the number of columns of the vector map M of the salient region; s represents a salient feature, the full set of S is {1,2, \8230;, n }, S = { i | ICL (p) _i )>0}；

and step B4: and converting the vector image M of the salient region into a two-dimensional image matrix M' serving as a salient matrix detection image.

The water surface target tracking method based on the significance detection further comprises a target extraction step;

the target extraction step comprises the following steps:

step C1: for the current frame, assuming that the target position of the previous frame is Pn-1, the current salient matrix detection graph M' totally has k high-brightness areas, marking the ith connected high-brightness area as Wi, obtaining the position coordinate Pi of the center point of Wi, and obtaining the maximum gray value PIX of the pixels in Wi _i And possesses a maximum gray value PIX _i The position information Mi of the pixel of (1); wherein i =1,2, \8230, k;

-defining a minimum number of pixels N _min If Wi has pixel number N _i Less than N _min If yes, filtering the Wi;

-defining a minimum grey value PIXmin if the maximum grey value PIX of Wi _i Less than PIX _min Then Wi is filtered;

and C3: filtering out invalid highlight areas through the step C2, and then filtering out the position coordinates Pi of the central points of the Wi and the distance D of the target position of the previous frame _i And (3) calculating:

selecting each D _i Minimum distance Dmin of (1);

In a preferred embodiment, the step A1 includes the following steps:

and constructing an image preprocessing Gaussian pyramid. The Gaussian pyramid construction process is divided into two parts, one part is to perform Gaussian blur of different scales on the image, the other part is to perform down sampling on the image, as shown in FIG. 1, the first layer is the original image, the size of the row and column is W x H, W and H are positive integers, from the second layer of image, the resolution of each layer of image is one fourth of the previous layer of image, and the resolution of the ith layer of image is (W/2i, H/2 i).

The Gaussian pyramid can effectively perform downsampling on the image, and the original information of the image is compressed with the highest quality. For example, the input original image resolution is 1920 × 1080, and in order to satisfy both the real-time performance requirement of the calculation and the calculation accuracy, the image resolution is reduced to 240 × 135 by using the 4 th layer gaussian pyramid image.

In a preferred embodiment, the step A2 includes the following steps:

and (4) color space conversion. Converting the color space of the image from the RGB space to the HSV space, wherein the specific conversion formula is as follows:

v＝max

where h denotes a hue component, max denotes the maximum of three components R, G, and B in the RGB color space, min denotes the minimum of three components R, G, and B in the RGB color space, R denotes a red component, G denotes a green component, B denotes a blue component, s denotes a saturation component, and v denotes a luminance component.

In the practical application process, the conversion of the color space can be realized by using the cvtColor function in the opencv library.

Setting a hue threshold value T _ h =50 and a saturation threshold value T _ s =30, and determining that the water surface or sky area is to be distinguished for an area in which a hue component h is smaller than the hue threshold value T _ h or a saturation component s is smaller than the saturation threshold value T _ s in the image. Setting a brightness threshold value T _ v =180, if the brightness component v of the water surface or sky area to be distinguished is smaller than the brightness threshold value T _ v, determining the water-land line area, and if the brightness component v of the water surface or sky area to be distinguished is larger than or equal to the brightness threshold value T _ v, determining the sky area. And obtaining a binary water bank line area image. The shoreline region is preferably a region comprising an onshore scene, a shoreline and a water surface reflection.

In a preferred embodiment, non-zero pixel points in the image are counted line by line, the number of the non-zero pixel points in the ith line is Ni, and if Ni is greater than 80, the image is judged to be a water shoreline region, so that the ROI is further accurate.

In a preferred example, the step of performing median filtering between step A2 and step A3 specifically includes:

and (5) median filtering processing. And D, filtering pulse noise of the binary water bank line region image obtained in the step A2 by using median filtering, wherein the formula is as follows:

g(x,y)＝median{f(x-i,y-j)}(i,j)∈W

wherein g represents the image after median filtering, g (x, y) represents the gray value of the pixel point at the (x, y) position, mean {. Represents taking the median, x represents the abscissa of the pixel point, y represents the ordinate of the pixel point, f represents the image before filtering, f (x-i, y-j) represents the gray value of the pixel point at the (x-i, y-j) position, W represents a two-dimensional template, preferably 3 × 3 templates are selected, i represents the abscissa of the two-dimensional template, and j represents the ordinate of the two-dimensional template.

And obtaining a filtered water bank line area image.

In a specific application process, median filtering can be performed by using a mediablend function in an opencv library.

In a preferred embodiment, the operator used for edge detection of the water bank line region image may be a Roberts Cross operator, a Prewitt operator, a Sobel operator, a Kirsch operator, a compass operator, a Canny operator, a Laplacian operator, or the like, and the following improved Canny edge detection is further used, specifically:

carrying out improved canny edge detection on the filtered water bank line area image, which specifically comprises the following steps:

traversing the water bank line region image by using a sobel operator to obtain the gradient direction and the gradient magnitude of the water bank line region image;

the sobel operator is as follows:

wherein s is _x Representing operators in the direction of the abscissa, s _y Representing operators in the ordinate direction.

The calculation formula of the gradient magnitude and gradient direction is as follows:

g _x ＝(m ₆ +2m ₇ +m ₈ )-(m ₀ +2m ₁ +m ₂ )

g _y ＝(m ₂ +2m ₅ +m ₈ )-(m ₀ +2m ₃ +m ₆ )

wherein G [ i, j]The gradient size of a pixel at the position of a water bank line area image (i, j) is represented, i represents the horizontal coordinate of the water bank line area image, and j represents the vertical coordinate of the water bank line area image; m is ₀ ～m ₈ Expressing the gray value of a pixel point at the position of the corresponding sobel operator of the water bank line region image; g _x Denotes the magnitude of the horizontal gradient, g _y And M represents a pixel point set in a3 x 3 region in the image.

Defining a weighting function

DoG _σ (x,y)＝g _4σ (x,y)-g _σ (x,y)

w _σ Representing a weight function, g (x, y) representing a two-dimensional Gaussian function, σ representing a scale parameter, | | ₁ Denotes the L1 norm, doG _σ Representing all of the DoG _σ Vector of (x, y) | H (DoG) _σ )‖ ₁ Representing all of the DoG _σ Sum of absolute values (x, y); g _σ (x,y)，g _4σ (x, y) represents a two-dimensional Gaussian function with scale parameters of sigma and 4 sigma, z represents a rational number, H (z) represents that when z is less than 0, z is greater than 0, and DoG _σ (x, y), subtracting two-dimensional Gaussian functions with the representation scale parameters of 4 sigma and sigma respectively, wherein x represents the horizontal coordinate of the pixel point, and y represents the vertical coordinate of the pixel point.

The method has the advantages that the edge detection can be carried out after the water shoreline area is obtained, the improvement of texture suppression can be carried out aiming at the characteristic that invalid edges are easy to detect by the edge detection, the suppression of similar textures is carried out in a circular area determined by a two-dimensional Gaussian function, and the judgment of the similar textures can be defined by utilizing the gradient size and the direction of the pixel point.

The above formula can be used to suppress the surrounding texture, and for the current pixel point, if the gradient size and direction of the current pixel point are similar to those of other pixel points in the neighborhood ring, the current pixel point is suppressed. And obtaining the water bank line image subjected to texture inhibition.

Where | is the L1 norm. Radii r1 and r2 are defined as shown in fig. 2. Within a circle of radius r1, doG _σ (x,y)&lt, 0, weight w thereof _σ Is 0, specifically σ =1,r ₁ ＝2σ，r ₂ ＝4r ₁ 。

In a preferred embodiment, the Canny () function in the opencv library is used for edge detection, the low threshold is taken as 100, and the high threshold is taken as 400.

In a preferred embodiment, the step of filtering the edge information is performed between the step A3 and the step A4, which is as follows:

7 operators are designed to traverse the water bank line image after texture suppression, and invalid edges of the image can be reduced.

Respectively as follows:

in a preferred embodiment, step A4 includes the following steps:

hough transform

ρ＝xcosθ+ysinθ

Wherein rho represents the geometric vertical distance from the original point to a straight line passing through the (x, y) point, and theta represents the included angle between the vertical line and the x axis; the method can convert pixel points of an image from (x, y) space to (rho, theta) space, for each pixel point (x, y) in the water bank line image after texture suppression, calculating the corresponding rho of theta from minus 180 degrees to 180 degrees at a determined angle interval delta theta, converting the rho to parameter space (rho, theta), voting and accumulating the number of the pixel points in the parameter space (rho, theta), and obtaining the longest straight line from the voting accumulated number. Considering that the absolute value of the included angle between the water bank line and the horizontal line is generally not more than 30 degrees, the detection range of Hough transform is adjusted, and the calculation range of theta is-30 degrees to 30 degrees.

In a preferred example, hough transformation is performed by HoughLines of Opnecv, and straight lines are extracted from the image to identify the water bank.

In a preferred embodiment, the step of detecting the significance includes the following steps:

firstly, a sparse dictionary is utilized to carry out sparse feature representation on an image, A is set as a sparse base, a _i Is the ith basis function. Let W be the filter function, W = A ^-1 The superscript-1 denotes inversion of the rarefaction radical A, W = [ W = ₁ ,w ₂ ,…,w ₁₉₂ ] ^T W of each row _j J =1,2, \ 8230;, 192, can be seen as a linear filter to the set of images to be processed.

The sparse representation s of the image can be seen as a linear filter w for all linear filters _j In response to (3). Given a vectorized representation x of an image, it can be expressed as s = Wx.

The incremental coding length is a significance expression method which accords with human visual characteristics and can enable the system energy distribution to be optimal.

Definition of p _i For the ith basis function X = [ X ] for sampling matrix ¹ ,x ² ,…,x ^k ,…]Wherein the sampling matrix X is a vectorized image matrix. Can give p _i Comprises the following steps:

wherein k is a positive integer and has a value range of [1,192]，w _i ∈W，x ^k e.X, i is a positive integer, and the value range of i is [0,192]。

The probability density function p for the basis functions with respect to the image response can be defined as: p = | p ₁ ,p ₂ ,…,p _n | ^T . Defining the information entropy of p as H (p), calculating discrete information entropyThe formula is as follows:

where H (p) is the information entropy and n represents the number of columns of the probability density function p.

The increment of the information entropy function is calculated as follows:

wherein: p is a radical of _j Represents the jth column of the probability density function p, j having a value range of [1, n ]]；p _i The ith column representing the probability density function p, i having a value in the range of [1, n ]]；

It follows that the delta code length ICL is:

wherein, ICL (p) _i ) Denotes with respect to p _i The delta code length of (c).

The full set of defining salient features S, S is {1,2, \8230;, n }, S = { i | ICL (p) _i )&gt, 0}. Defining the energy obtained by the ith significant feature S as d _i If the current feature S ₀ Not of a salient feature, then

Wherein, d _k Is not shown as belonging toThe characteristic of the obvious characteristic, k is a positive integer, and the value range is the complement of S;

finally, one image X = [ X ] is given ¹ ,x ² ,…,x ⁿ ]The significant region vector diagram M, M = [ M ] can be obtained through quantification ₁ ,m ₂ ,…,m _n ]Wherein:

wherein m is _k The k column of the vector diagram representing the salient region, wherein the value range of k is [1, n ]]。

In order to convert the salient region map into an image for display, a normalization process is required. First, for the salient region map vector M, the absolute value of each element is determined, and then converted to the [0,1 \8230;, 255] domain, and then the vectorized salient region map is converted to a two-dimensional image matrix M', which is the final salient matrix detection map.

More specifically:

in the first step, the ROI is taken as a processing image, and is normalized into a double type within the range of 0-1, namely, the linear normalization of a 0-255int type is carried out to the 0-1double type.

And secondly, converting the two-dimensional image into a vector form. The transformation mode is as follows: and (3) traversing the image from top to bottom and from left to right by using an 8 × 8 operator to fetch points, and filling the 8 × 8 matrix fetched every time into the vector from left to right and from top to bottom.

Thirdly, performing sparse feature representation, and setting A as a sparse base and ai as i ^th A basis function. Let W = A ^-1 For the filter function, W = [ W = ₁ ,w ₂ ,…,w ₁₉₂ ] ^T W of each row _j Can be viewed as a linear filter on a collection of images. The sparse representation s of the image can be seen as a linear filter w for all linear filters _j In response to (3). Given a vectorized representation x of an image, it can be expressed as s = W x. Matrix multiplication is carried out on the image converted into the one-dimensional vector to obtain probability densityThe function p.

And the GPU parallel operation is utilized to improve the efficiency of matrix multiplication and save the calculation time. The method specifically comprises the following steps of installing a Cublas basic linear subprogram library by utilizing a CUDA parallel computing architecture.

First, calling cublascredit () function to initialize Cublas. And then performing matrix multiplication on the matrix defined in the memory and the image vector by using a cudaMemcpy () function. And finally, inputting the two matrixes by adopting a cudaMalloc () function, and calculating to obtain a probability density function.

And fourthly, calculating the incremental coding length.

And fifthly, calculating energy. Defining salient features S, S = { i | ICL (p) _i )&gt 0} defining the energy obtained by each salient feature as d _i If the feature does not belong to a salient feature, then define

And sixthly, performing matrixing on the vector to convert the vector into a two-dimensional image.

Seventhly, an image X = [ X ] is given ¹ ,x ² ,…,x ⁿ ]The significant region map M = [ M ] can be quantized ₁ ,m ₂ ,…,m _n ]Wherein:

in a preferred embodiment, the information extracting step includes the following steps:

the method comprises the steps that information extraction needs to be carried out on a significant matrix detection map M', the target position of the previous frame is assumed to be Pn-1, k highlight areas are totally arranged in the current significant matrix detection map, each communicated highlight area is marked as Wi, pixel-by-pixel statistics and analysis are carried out on each Wi, the number Ni of pixels of the highlight areas is obtained, the position information of the edge pixels of Wi is obtained, therefore, the position coordinate Pi of the center point of the current highlight area is obtained, and in addition, the position information Mi of the pixel with the maximum gray value and the maximum gray value PIXi are counted.

Analyzing the highlight area Wi, and setting the following filtering conditions: defining the minimum number of pixels Nmin if the number of pixels N of Wi _i <N _min Filtering it; defining the minimum gray value PIXmin if the maximum gray value PIX of Wi _i <PIX _min Filtering it; wi is polygon approximated and filtered if the approximation is a concave polygon.

After filtering out invalid highlight area information, performing distance calculation on the center point position Pi of the Wi left after filtering and the target position of the previous frame:and selecting the minimum distance Dmin, if the minimum distance Dmin is smaller than a set maximum distance threshold DT, determining a highlight area Wi which currently has the minimum distance with a target of a previous frame as a target area, and if the minimum distance Dmin is larger than the maximum distance threshold DT, determining that the target of the current frame is lost, and adopting the position Pn-1 of the previous frame as the position of the current target. The specific flow is shown in fig. 6.

The foregoing description has described specific embodiments of the present invention. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes or modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention. The embodiments and features of the embodiments of the present application may be combined with each other arbitrarily without conflict.

Claims

1. A water surface target tracking method based on significance detection is characterized by comprising a water bank line detection step;

the water shoreline detection step comprises the following steps:

step A1: down-sampling the image to obtain a down-sampled image;

step A3: carrying out edge detection on the water bank line region image;

step A4: carrying out Hough transform on the detected edge to obtain a water shoreline;

the significance detection step comprises the following steps:

ICL(p _i )＝-H(p)-p _i -logp _i -p _i logp _i

wherein, ICL (p) _i ) Denotes with respect to p _i The delta code length of (d); p is a radical of _i The ith row of the probability density function p is represented, i having a value range of [1, n ]](ii) a The probability density function p is defined as: p = | p ₁ ，p ₂ ，...，p _n | ^T ；

wherein k is a positive integer and has a value range of [1, 192%]；w _i The value range of the epsilon W, i is [0,192](ii) a W is the filter function, W = A ^-1 A is the sparse basis of the ROI of interest, W = [ W = ₁ ，w ₂ ，…，w ₁₉₂ ] ^T ；

x ^k E.x, X denotes the sampling matrix, X = [ X ] ¹ ，x ² ，...，x ^k ，...](ii) a The sampling matrix X is an image matrix of the region of interest ROI subjected to vectorization;

M＝[m ₁ ，m ₂ ，...，m _n ]

wherein m is _k The k column of the vector diagram M representing the salient region, k having a value range of [1, n]N is the number of columns of the vector map M of the salient region; s represents a salient feature, the full set of S is {1,2, \8230;, n }, S = { i | ICL (p) _i )＞0}；

Wherein p is _j The j-th column of the probability density function p is represented, and the value range of j is [1, n ]]；ICL(p _j ) Denotes with respect to p _j The delta code length of (d);

and step B4: converting the vector image M of the salient region into a two-dimensional image matrix M' serving as a salient matrix detection image;

the target extraction step comprises the following steps:

step C1: for the current frame, assuming that the target position of the previous frame is Pn-1, the current salient matrix detection map M' totally has k high-brightness areas, marking the ith connected high-brightness area as Wi, obtaining the position coordinate Pi of the center point of the Wi, and obtaining the maximum gray value PIX of the pixel in the Wi _i And possesses a maximum gray value PIX _i The position information Mi of the pixel of (1); wherein i =1,2, \8230, k;

and C2: filtering each connected highlight area Wi to filter out invalid highlight areas; the following filtration conditions were set:

wherein x is _n-1 Abscissa, y, representing the position of the target in the previous frame _n-1 Ordinate, x, representing the position of the object in the previous frame _i Abscissa, y, representing the location of the center point of Wi _i A vertical coordinate representing a position of a center point of Wi;

selecting each D _i Minimum distance Dmin of (1);

2. A water surface target tracking method based on saliency detection as claimed in claim 1 characterized by hue threshold T _ h =50, saturation threshold T _ s =30, brightness threshold T _ v =180.