Method for rapidly tracking and positioning set targets in grayscale videos

A gray-scale video, tracking and positioning technology, applied in image data processing, instruments, calculations, etc., can solve the problems of complex tracking process, increased calculation amount and complexity, slowness, etc., and achieve high tracking accuracy, fast calculation speed, Effects of adaptation to scale changes

Inactive Publication Date: 2018-06-29
XIDIAN UNIV
0 Cites 12 Cited by

AI-Extracted Technical Summary

Problems solved by technology

[0003] However, none of the existing tracking methods can directly adapt to the scene where the target scale changes. The fixed scale makes it not only unable to accurately output the coordinate position of the tracked target in the scene where the target scale changes, but also unable to calibrate the target ...
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Abstract

The invention relates to a method for rapidly tracking and positioning set targets in grayscale videos. The method comprises the following steps of: carrying out two-dimensional Gaussian kernel cyclicconvolution on a standard target image and an image in a video frame sequence to solve cross-correlation matrixes, and processing positions of targets tracked by the cross-correlation matrixes by using ridge regression method; carrying out weighted average on a detection result of the current frame to serve as a next frame of standard target image; self-convoluting each frame of standard target image and solving a statistic feature of each frame of standard target image, and mapping characterization between two frames as scale change so as to solve scale change, in a moving process, of the target; and deconvolution is carried out on an auto-correlation matrix through discrete Fourier transformation and inverse transformation to update and learn a ridge regression parameter which is used for solving a response matrix in the next frame. The method is high in calculation speed and high in tracking precision, adopts dense sampling, is completely adapted to scale change and can be suitablefor the scenes in which targets are moved from far to near or moved from near to far.

Application Domain

Technology Topic

Image

  • Method for rapidly tracking and positioning set targets in grayscale videos
  • Method for rapidly tracking and positioning set targets in grayscale videos
  • Method for rapidly tracking and positioning set targets in grayscale videos

Examples

  • Experimental program(1)

Example Embodiment

[0051] In order to further illustrate the technical means and effects of this embodiment to achieve the predetermined purpose, the specific implementation, structural features and effects of this embodiment will be described in detail below in conjunction with the drawings and embodiments.
[0052] The technical solution adopted in this embodiment is mainly divided into four parts: a preprocessing phase, a tracking phase, a parameter learning phase, and a scale prediction phase. In the first frame, it is necessary to artificially specify the position and size of the target in the first frame, and then crop the original image to obtain a standard image centered on the tracking target.
[0053] Step 1: Obtain the current frame tracking window based on the target coordinates and size calculated in the previous frame, and perform logarithmic transformation and cosine window smoothing on the tracking window and the standard target image.
[0054] Step 2: The preprocessed tracking window is cyclically convolved with the standard target image to obtain a cross-correlation matrix.
[0055] Step 3: Perform ridge regression on the cross-correlation matrix to obtain the coordinates of the target in the current frame.
[0056] Step 4. Re-update the target standard map according to the newly obtained coordinates.
[0057] Step 5: Calculate the standard deviation of the autocorrelation matrix of the current standard graph, quoting the standard deviation of the autocorrelation matrix of the current frame and the previous frame standard to obtain the proportional coefficient P, and update the target scale.
[0058] Step 6. Use the discrete Fourier transform and its inverse transform to perform deconvolution of the standard response graph for the autocorrelation matrix, and update the learning parameters of the ridge regression.
[0059] In this embodiment, x (i, j) and z (i, j) are used to represent the i-th row and j-th column of the matrix, and x i And z i Represents the matrix of the i-th frame.
[0060] The detailed process of this embodiment is:
[0061] 1. Pretreatment stage:
[0062] In order to enhance the contrast of the image and make the image have a more obvious contrast in the darker area, the video frame is firstly transformed by logarithm:
[0063] x(i,j)=clog(1+x(i,j))
[0064] Where c is the constant coefficient of logarithmic transformation, and x(i,j) is the pixel value of the corresponding coordinate of the single frame picture. At the same time, in order to eliminate asymmetric noise, make the target in the center of the single frame image more prominent and eliminate noise, then use the cosine window to multiply the single frame image, and smoothly process the target image with a cosine window consistent with the image length and width. Assuming that the height of each frame of image is h and the width is l, the method of constructing the cosine window consists of two sizes of 1×h and 1×l cosine vector V h And V l The cross product of to obtain a matrix W of size h×l:
[0065] W=V h T ×V l
[0066]
[0067]
[0068] W is the constructed cosine window, and the image after logarithmic transformation is smoothed with the W matrix: x=x·W.
[0069] 2. Tracking stage:
[0070] The target is tracked by means of cross-correlation filtering, and the coordinate position of the target in the current frame is calculated. According to the target coordinate position and target size calculated in the previous frame, the current frame to be detected image x is obtained by cropping. Calculate the cross-correlation matrix between the image x to be detected and the standard image z of the previous frame by Gaussian kernel circular convolution:
[0071] Do cyclic matrix C for the current frame to be detected x(i,j) Two permutation matrices T are available i With T j To represent:
[0072] C x(i,j) = T i xT j
[0073] T in the above formula i It is the unit array row operation, which is obtained by cyclic displacement i times, the same way T j It is a unit array operation, which is obtained by cyclic shift j times. C x(i,j) That is, the cyclic displacement matrix when the image x to be convolved with the size of m×n is in the i-th row and j-th column. In this embodiment, the Gaussian kernel method is used to perform cyclic convolution on the standard image of the previous frame and the image to be detected, thereby obtaining the cross-correlation matrix:
[0074]
[0075] Is the cross-correlation matrix The value of the element in row i and column j. Each step in the formula is calculating the image x to be convolved and its cyclic displacement matrix C at row i and column j x(i,j) The correlation between the two, the more similar the two, the correlation (ie ) Is higher, β is the Gaussian kernel bandwidth. .
[0076] After the cross-correlation matrix is ​​obtained, the ridge regression method is used to learn and locate the target position. The objective function of the ridge regression learning function classifier is:
[0077] Parameter α i With The coefficient matrix with the same length and width is the learning parameter of the ridge regression method. Convolve the two to get R(x i ). R(x i ) Is the i-th frame image x i The cross-correlation response matrix with the standard target image z. The coordinate position of the tracking target is at the highest peak of the response matrix.
[0078] 3. Parameter learning stage:
[0079] After the tracking stage, the target coordinate position is obtained, and the image is cropped with the new coordinate as the center to obtain the new standard image z' i+1. The new and old standard graphs are updated by coefficient weighted average to obtain the standard graph used in the next frame calculation, where θ is the weighting coefficient:
[0080] z i+1 =θz' i+1 +(1-θ)z i
[0081] In this embodiment, the learning parameters are updated and the target scale is calculated by calculating the autocorrelation matrix of the standard graph. The method of calculating the autocorrelation matrix is ​​the same as that of the cross-correlation matrix:
[0082]
[0083] The update method of the learning parameters of the ridge regression method is:
[0084]
[0085] Where y is the standard response matrix, which is constructed in this embodiment as a Gaussian model centered on the geometric midpoint of the image and equal to the length and width of the image. Through the discrete Fourier transform of the standard response matrix and the standard graph autocorrelation matrix, the inverse discrete Fourier transform is performed after the matrix points are divided in the frequency domain to quickly realize the deconvolution process, and obtain the next frame of ridge regression learning parameters .
[0086] Fourth, the scale calculation stage
[0087] The change of the autocorrelation matrix of the target image can reflect the change rule of the target scale in the center of the image. Respectively represent the autocorrelation matrix of the previous frame and the next frame, and s represents the target scale, then:
[0088]
[0089] Where g is about Function mapping, the proportional coefficient P represents from To The scale changes, there is s A =P·s B , The process of finding the current scale of the target is the process of finding the proportional coefficient P. Because there are too many elements in the matrix, the processing of the matrix directly is more complicated, and the pertinence is not strong. If the statistics of the autocorrelation matrix are used as the independent variables of the mapping function, the amount of calculation and complexity will be well optimized.
[0090] The autocorrelation matrix is ​​obtained by Gaussian kernel convolution, so the result shows a shape similar to a two-dimensional Gaussian model. The most important statistical parameter of the Gaussian model is the standard deviation σ. The size of σ determines the degree of central distribution of the Gaussian model. The larger the σ, the more dispersed the Gaussian distribution. In scale detection, the target is larger and its autocorrelation matrix is ​​more scattered. It can be seen that σ is positively correlated with the target scale s. Therefore, this paper uses the standard deviation σ of the image autocorrelation matrix as the independent variable of the mapping function g to calculate the scale factor P. The calculation formula of σ is as follows:
[0091]
[0092] Where N is the number of pixels, Is the value of the i-th row and j-th column of the autocorrelation matrix, and u is the matrix mean. The standard deviation is used as the independent variable. Due to the requirement to solve the proportional relationship, this embodiment directly uses σ B And σ A As the argument of the mapping function, take the mapping function g as:
[0093]
[0094] So far, the current scale of the target can be solved by the scale factor P. Single frame tracking and scale calculation are over.
[0095] This embodiment has fast calculation speed, adopts dense sampling; high tracking accuracy; can fully adapt to scale changes, and can be applied to scenes where the target quickly moves from far to near or from near to far.
[0096] The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be considered that the specific implementation of the present invention is limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, a number of simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention.
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

no PUM

Description & Claims & Application Information

We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Similar technology patents

Classification and recommendation of technical efficacy words

  • Calculation speed
  • Improve tracking accuracy

Log correlation analysis system and method

ActiveCN101610174ACalculation speedReduce false alarmsData switching networksEvent correlationCross correlation analysis
Owner:SHENZHEN Y& D ELECTRONICS CO LTD

Face detection method and device

InactiveCN107871134ACalculation speedThe classification prediction is accurateCharacter and pattern recognitionFalse detectionNetwork model
Owner:BEIJING EYECOOL TECH CO LTD

Facial tracking method

Owner:PRIMAX ELECTRONICS LTD
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products