CN103077530A - Moving object detection method based on improved mixing gauss and image cutting - Google Patents
Moving object detection method based on improved mixing gauss and image cutting Download PDFInfo
- Publication number
- CN103077530A CN103077530A CN2012103659948A CN201210365994A CN103077530A CN 103077530 A CN103077530 A CN 103077530A CN 2012103659948 A CN2012103659948 A CN 2012103659948A CN 201210365994 A CN201210365994 A CN 201210365994A CN 103077530 A CN103077530 A CN 103077530A
- Authority
- CN
- China
- Prior art keywords
- target
- frame
- num
- image sequence
- field picture
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Abstract
The invention relates to a moving object detection method based on improved mixing gauss and image cutting, which belongs to the technical field of intelligent video monitoring, and aims to solve the problems of the existing method and improve the accuracy and the processing speed of moving object detection. The method comprises the following steps of: (1) confirming a moving area, wherein the invention provides a three-frame difference and curve fitting method based on gray level difference so as to preliminarily confirm the moving area of an object; (2) initializing an expectation-maximization (EM) algorithm, and confirming an initial value of the EM algorithm by utilizing a method based on grid and density estimation so as to reduce the dependency of the EM algorithm on the initial value; and (3) detecting the moving object, utilizing the initialized EM algorithm to estimate a mixing gauss model parameter, and only adopting a mixing gauss model method to detect the moving object on a cut moving area, so that the computation complexity is greatly reduced. According to the moving object detection method based on improved mixing gauss and image cutting, the accuracy of object detection can be ensured, and simultaneously, the algorithm can meet the performance requirement of actual application.
Description
Technical field
The invention belongs to the intelligent video monitoring field, be specifically related to a kind of moving target detecting method based on improving mixed Gaussian and image cut, be used for the moving object detection of video monitoring.
Background technology
Along with the fast development of artificial intelligence technology, the research of intelligent video monitoring system and application become the focus that people pay close attention to gradually, and are widely used in a plurality of fields such as safety guarantee, traffic, business activity and military affairs.In the scene of video monitoring, moving target is main body, so supervisory system must detect moving target in real time.Moving object detection is the image Segmentation Technology of how much of a kind of based targets and statistical nature, and main method has frame differential method, optical flow method, background subtraction point-score etc.
Frame differential method is a kind of simple directly method for testing motion, it utilizes target and the difference of background on gamma characteristic that will extract in the image, by choosing suitable gray threshold to Image Segmentation Using, thereby target is extracted from background, because it realizes simple, fast operation, in most of the cases detect effect better, usually be used for the occasion higher to requirement of real-time.But frame differential method is processed the gained image and is often contained a large amount of noises, hollow and division part, and effect is unsatisfactory when processing than the complicated image sequence.The ultimate principle that optical flow method detects moving object is to give a velocity to each pixel in the image, this has just formed an image motion field, in a particular moment of motion, point on the image is corresponding one by one with point on the three-dimensional body, this corresponding relation can be obtained by projection relation, according to the velocity feature of each pixel, can carry out performance analysis to image.The advantage of optical flow method is that light stream not only carried the movable information of moving object, but also has carried the abundant information of relevant scenery three-dimensional structure, and it can in the situation of any information of not knowing scene, detect Moving Objects.But optical flow method is calculated height consuming time, and real-time and practicality are all relatively poor.The background subtraction point-score is to utilize the difference of present image in the image sequence and background image to detect a kind of technology of moving region, but the variation for dynamic scene, responsive especially such as the interference of illumination and external extraneous events etc., only be fit to the little situation of environmental change.
For real-time and accurately inspected object variation, Wren(Pfinder:Real-time tracking of the human body.IEEE Trans.Patt. in 1997) propose to set up color model, this model Gaussian distributed for each pixel in the image.But single Gauss model is unsatisfactory to the treatment effect of outdoor environment.Stauffer in 1999 and Grimson(Adaptive background mixture models for real-time tracking, CVPR, 1999.IEEE Computer Society Conference on.) proposed the algorithm of mixed Gauss model, and be widely used in the complex scene background modeling of robust, but a great problem that faces of mixed Gauss model is exactly that the calculated amount of algorithm is large in actual applications, and real-time is relatively poor.In recent years, many researchers propose diverse ways and solve the poor problem of mixed Gauss model real-time.Wherein, expectation maximization (EM) algorithm is the classical way of mixed Gauss model parameter estimation.But the EM algorithm is relatively more responsive to the input initial value, and often difference is larger for cluster result corresponding to different initial values.
Summary of the invention
The present invention is directed to the problem that calculated amount is large, real-time is poor that mixed Gauss model shows aspect moving object detection, propose a kind of moving target detecting method based on improving mixed Gaussian and image cut, concrete experimental program is as follows:
The video that obtains for the fixing rig camera of visual field, at first utilize the three-frame difference of intensity-based otherness and the method for curve to determine the target travel zone, then in the scope of moving region, adopt the mixed Gauss model that merges the EM algorithm to carry out moving object detection, can reduce greatly calculated amount like this, shorten Riming time of algorithm; Wherein, utilization of the present invention is carried out initialization based on the method for grid and density Estimation to the EM algorithm, to reduce the EM algorithm for the dependence of initial value.
Specific implementation step of the present invention is:
(1) image input step, input raw video image sequence;
(2) image pre-treatment step converts the raw video image sequence to grayscale image sequence;
(3) image difference calculation step:
(3.1) get size and be the moving window W of M * N, W
iThe gray matrix of moving window in the i two field picture of (x, y) expression grayscale image sequence, the gray difference property coefficient R (W of the corresponding window of adjacent two frames in the calculating grayscale image sequence
i(x, y), W
I+1(x, y)), computing method are as follows:
Wherein, x and y are respectively horizontal ordinate and the ordinate of each element in the gray matrix, and M and N are respectively the length of moving window and wide, and 0＜R (W
i(x, y), W
I+1(x, y))＜1;
(3.2) if the gray difference property coefficient greater than threshold value T, then the gray-scale value with moving window correspondence position in the i two field picture of grayscale image sequence all is set as 1, otherwise be set as 0, the span of T is 0.25 to 0.35, mobile moving window is until i frame and the i+1 two field picture of traversal grayscale image sequence, repeat above operation until grayscale image sequence is all finished the computing of the gray scale difference opposite sex, obtain bianry image sequence I
B(x, y);
(3.3) to bianry image sequence I
BAdjacent two frames are ' with ' computing, bianry image sequence I in (x, y)
BAll images obtain target image sequence I after finishing ' with ' computing in (x, y)
M(x, y), I
MGray-scale value is that 1 zone is the preliminary moving target of determining in (x, y);
(4) moving region determining step:
(4.1) according to target image sequence I
MTarget location in (x, y) is determined specifically to comprise the direction of motion of target,
Target image sequence I
M(x, y) comprises the num two field picture altogether, if num is even number, from target image sequence I
MExtract the 1st frame, num/2 two field picture and num two field picture in (x, y), if num is odd number, then from target image sequence I
M(x, y) extract the 1st frame, (num+1)/2 two field picture and num two field picture in, and the peak of target location in 3 two field pictures carried out fitting a straight line, if fitting a straight line and horizontal direction angle are then judged target along continuous straight runs motion less than 45 °, if fitting a straight line and horizontal direction angle judge then that greater than 45 ° target vertically moves;
If target image sequence I
M(x, y) target location in the num two field picture appears at the left side of target location in (num+1)/2 frame or the num/2 frame, the target location appears at the left side of target location in the 1st frame in (num+1)/2 frame or the num/2 frame, judges that then target is to move from right to left; If the target location in the num two field picture appears at the right side of target location in (num+1)/2 frame or the num/2 frame, the target location appears at the right side of target location in the 1st frame in (num+1)/2 frame or the num/2 frame, judges that then target is to move from left to right; Target location in the num two field picture appears at the top of target location in (num+1)/2 frame or the num/2 frame, the target location appears at the top of target location in the 1st frame in (num+1)/2 frame or the num/2 frame, judges that then target is to move from bottom to top; If the target location in the num two field picture appears at the below of target location in (num+1)/2 frame or the num/2 frame, the target location appears at the below of target location in the 1st frame in (num+1)/2 frame or the num/2 frame,, judge that then target is to move from the top down;
(4.2) according to the direction of motion of target, determine object boundary point position, and utilize the method for least square curve fitting to determine the moving region of target;
When the target along continuous straight runs moves, with target image sequence I
M(x, y) peak of target location utilizes least square curve fit to obtain the upper bound in target travel zone in each frame, minimum point utilizes least square curve fit to obtain the lower bound in target travel zone, if target is from right to left motion, the rightest some place vertical line of the first frame, the upper bound lower bound that the most left some place vertical line of last frame and match obtain consists of the moving region of target, if target is from left to right motion, the most left some place vertical line of the first frame, the upper bound lower bound that the rightest some place vertical line of last frame and match obtain consists of the moving region of target; When target is vertically moved, with target image sequence I
M(x, y) the most left point of target location utilizes least square curve fit to obtain the left margin in target travel zone in each frame, the rightest point utilizes least square curve fit to obtain the right margin in target travel zone, if target is from the top down motion, the first frame peak place horizontal line, the upper bound lower bound that last frame minimum point place horizontal line and match obtain consists of the moving region of target, if target is from bottom to top motion, the first frame minimum point place horizontal line, the upper bound lower bound that last frame peak place horizontal line and match obtain consists of the moving region of target;
(4.3) with in each frame of grayscale image sequence corresponding to the Partial Shear of moving region out, consist of moving region image sequence I
q(x, y);
(5) expectation maximization EM algorithm initialization step:
(5.1) moving region image sequence I
qFirst two field picture of (x, y) is as the input data of expectation maximization EM algorithm initialization step, moving region image sequence I
qFirst two field picture of (x, y) is the data space of a L * H, and wherein L is moving region image sequence I
qFirst two field picture of (x, y) shared number of pixels on the x direction of principal axis, H is moving region image sequence I
qFirst two field picture of (x, y) shared number of pixels on the y direction of principal axis;
Read 1 data point A from the input data of EM algorithm initialization step, coordinate is (x, y), and sets up the rectangular node centered by A, and the rectangular node method for building up is as follows:
Rectangular node is centered by A, and long is E
Ps1, wide is E
Ps2The zone, that is:
Wherein,
K is the mixed Gauss model number of components;
(5.2) reading data point successively from the input data of EM algorithm initialization step, if this point can not fall into any one existing grid, then according to the newly-built grid of the method for step (5.1), the density of grid is 1, if the data point of new input belongs to existing grid, corresponding mesh-density increases by 1, until data point is all handled in the input data;
If mesh-density is high fine grid with this grid mark then greater than threshold value den, the value of den is 20% of input data amount check;
(5.3) have the public domain when two high fine grids, perhaps two high fine grids all follow another high fine grid to be communicated with, and think that namely these two high fine grids are interconnected; The highly dense grid cell that is communicated with is combined as high fine grid family, and with the center of gravity of each high fine grid family as initial cluster center;
(5.4) utilize k average (k-means) algorithm that the input data of EM algorithm initialization step are gathered and be C class, the size of C equals the number of initial cluster center in the step (5.3), processes cluster result as the initial parameter value of EM algorithm
∑ and μ, wherein, weights Φ equals the ratio of total strong point number of the input data of data point number in each class and EM algorithm initialization step, and the variance ∑ equals the covariance of data point in each class, and average μ equals the center of each class;
(6) utilize initialized EM algorithm to estimate mixed Gauss model GMM parameter, then utilize mixed Gauss model to detect moving target, wherein the input data of mixed Gauss model are moving region image sequence I
q(x, y).
Beneficial effect:
⑴ the present invention adopts the three-frame difference of intensity-based otherness and the method for curve to determine the target travel zone, then out make it consist of the moving region image sequence corresponding to the Partial Shear in target travel zone each frame in the original input gray level image sequence, recycle improved mixed Gauss model and in the scope in target travel zone, carry out moving object detection.The inventive method had added target travel zone determining step before utilizing mixed Gauss model detection moving target, greatly dwindled the calculated amount of mixed Gauss model, shorten the operation time of algorithm of target detection, and then improved the real-time of algorithm of target detection.
⑵ the present invention adopts the method initialization EM algorithm based on grid and density Estimation, has reduced the dependence of EM algorithm for initial value.The method utilizes dynamic grid to reduce grid number, utilize the concept of density Estimation to abate the noise and isolated point data and be communicated with close quarters, generate initial cluster center by calculating the center of gravity that is communicated with close quarters, and to the input data carry out cluster, process cluster result and with its initial value as the EM algorithm, the more effective minimizing computing iterations of energy obtains comparatively desirable precision simultaneously, thereby has guaranteed the accuracy of target detection.
Description of drawings
Fig. 1 is schematic flow sheet of the present invention;
Fig. 2 (a)-2(f) is the parts of images of pedestrian's video sequence;
Fig. 3 (a)-3(f) is the parts of images of automobile video frequency sequence;
Fig. 4 is the coboundary, moving region that the vehicle sequence utilizes least square curve fit to obtain;
Fig. 5 (a), Fig. 5 (b) are respectively that the 1st frame, the 120th frame of pedestrian's sequence of video images is the last frame of pedestrian's sequence of video images;
White portion is the target travel zone of determining based on the image cut method that the present invention proposes among Fig. 5 (c), and black part is divided into background;
Fig. 5 (d) is based on the width of cloth in the pedestrian moving region image sequence that image cut method that the present invention proposes shears out;
Fig. 6 (a), Fig. 6 (b) are respectively that the 1st frame, the 60th frame of automobile video frequency image sequence is the last frame of automobile video frequency image sequence;
White portion is the target travel zone of determining based on the image cut method that the present invention proposes among Fig. 6 (c), and black part is divided into background;
Fig. 6 (d) is based on the width of cloth in the vehicle movement area image sequence that image cut method that the present invention proposes shears out;
Fig. 7 is that the EM algorithm compares based on the random value initialization with based on grid and the initialized convergence of density estimation method (GEEM) that the present invention proposes;
Fig. 8 (a), Fig. 8 (b), Fig. 8 (c) are the 1st frame, the 50th frame and the 100th frames of pedestrian's sequence of video images;
Fig. 8 (d), Fig. 8 (e), Fig. 8 (f) are that pedestrian's sequence of video images is based on the result of Traditional GM M algorithm moving object detection;
Fig. 8 (g), Fig. 8 (h), Fig. 8 (i) are that pedestrian's sequence of video images is based on the result of algorithm of the present invention (CIGMM) moving object detection;
Fig. 9 is conventional hybrid Gauss model (GMM), merges the method (EM-GMM) of random initializtion EM and mixed Gauss model and the comparison consuming time that put forward the methods of the present invention (CIGMM) detects moving target.
Embodiment
The present invention utilizes the three-frame difference of the gray scale difference opposite sex and the method for curve to determine the target travel zone; Then, utilization is based on the method initialization EM algorithm of grid and density Estimation; At last, use initialized EM algorithm to estimate the mixed Gauss model parameter, thereby can obtain fast and accurately the moving target foreground information.Treatment scheme is as shown in Figure 1:
⑴ input original video, and utilize that the avi2im function splits into continuous image sequence with video among the Matlab, and it is avi that the avi2im function requires the input video form, if the experiment video is extended formatting, can be converted into the avi form by winavi software.Fig. 2 (a)-2 (f), Fig. 3 (a)-3 (f) are respectively the present invention and test the pedestrian who adopts, the parts of images of automobile video frequency sequence.
⑵ image pre-treatment step utilizes that the rgb2gray function converts the raw video image sequence of inputting in the previous step to grayscale image sequence among the Matlab.
⑶ get 5 * 5 moving window W, W from grayscale image sequence the 1st frame
1The gray matrix of moving window W in the 1st two field picture of (x, y) expression grayscale image sequence, W
2(x, y) gray matrix of moving window W in the 2nd two field picture of expression grayscale image sequence, moving window W is mobile in gray level image the 1st frame and the 2nd frame synchronously, the gray difference property coefficient R (W of the moving window W of correspondence position in the 1st frame and the 2nd frame in the calculating grayscale image sequence
i(x, y), W
I+1(x, y)), computing method are as follows:
Wherein, x and y are respectively horizontal ordinate and the ordinate of each element in the gray matrix, M=N=5, and 0＜R (W
i, W
I+1)＜1, then,
Wherein, I
Bi(x, y) expression bianry image sequence I
BThe i two field picture of (x, y), T gets 0.25, as otherness coefficients R (W
i, W
I+1) then assert it is target during greater than T, on the contrary be background.The not overlapping mobile moving window of level is until reach in the horizontal direction last pixel of gray level image, then moving window is vertically moved down 5 pixels, the not overlapping mobile moving window of continuation level repeats above step until the 1st frame and the 2nd two field picture of traversal grayscale image sequence.If gray level image in the horizontal direction or on the vertical direction shared pixel can not be divided exactly then by 5 that last window comprises the residual pixel that the party makes progress.
Continuation repeats above operation to the 2nd frame and the 3rd two field picture of grayscale image sequence, until grayscale image sequence is all finished the computing of the gray scale difference opposite sex, namely until the last frame of grayscale image sequence and two field picture second from the bottom are repeated above operation, so far obtain bianry image sequence I
B(x, y).With bianry image sequence I
BAdjacent two frames of (x, y) are done ' with ' computing, with the noise of further elimination difference image.
I
Mi(x，y)＝I
Bi(x，y)&I
B(i+1)(x，y)
Wherein, I
Mi(x, y) expression target image sequence I
MThe i two field picture of (x, y).
Bianry image sequence I
BAfter (x, y) all finishes computing, obtain target image sequence I
M(x, y), I
MGray-scale value is that 1 zone is the preliminary moving target of determining in (x, y).
⑷ from the target image sequence I of pedestrian's video
M(x, y) extract the 1st frame in, the 60th frame and the 120th frame, the peak of target location in the 1st frame, the 60th frame and the 120th frame utilized the ployfit function carries out fitting a straight line among the Matlab, match gained straight line and horizontal direction angle are 0 °, and the target in the 60th two field picture appears at the left side of the 1st frame, and the target in the 120th two field picture appears at the left side of the 60th frame, and the target of then judging pedestrian's video sequence is that along continuous straight runs moves from right to left; Similarly, from the target image sequence I of automobile video frequency
M(x, y) extract the 1st frame in, the 30th frame and the 60th frame, the peak of target location in the 1st frame, the 30th frame and the 60th frame utilized the ployfit function carries out fitting a straight line among the Matlab, match gained straight line and horizontal direction angle are 10 °, and the target in the 30th two field picture appears at the right of the 1st frame, and the target in the 60th two field picture appears at the right of the 30th frame, and the target of then judging the automobile video frequency sequence is that along continuous straight runs moves from left to right.
According to the direction of motion of target, determine object boundary point position, and utilize the method for least square curve fitting to determine the moving region of target.Target is from right to left along continuous straight runs motion in pedestrian's video sequence, with target image sequence I
M(x, y) peak of target location utilizes least square curve fit to obtain the upper bound in target travel zone in each frame, the minimum point of target location utilizes least square curve fit to obtain the lower bound in target travel zone, the rightest some place vertical line of the first frame target location, the upper bound lower bound that the most left some place vertical line of last frame target location and match obtain consists of the moving region of target; Similarly, target is that along continuous straight runs moves from left to right in the automobile video frequency sequence, with target image sequence I
M(x, y) peak of target location utilizes least square curve fit to obtain the upper bound in target travel zone in each frame, the most left some place vertical line in the first frame target location, the upper bound lower bound that the rightest some place vertical line in last frame target location and match obtain consists of the moving region of target.With in each frame of grayscale image sequence corresponding to the Partial Shear of moving region out, consist of moving region image sequence I
q(x, y).Fig. 4 is the coboundary, moving region that the vehicle sequence utilizes least square curve fit to obtain.Fig. 5 determines the target travel zone for three-frame difference and the curve fitting method of pedestrian's video sequence intensity-based otherness.Fig. 5 (a) is the 1st frame of pedestrian's video image, Fig. 5 (b) is the last frame of pedestrian's video image, i.e. the 120th frame, dotted line among Fig. 5 (a), Fig. 5 (b) represents the border of target, the direction of motion that arrow represents target, white portion is the target travel zone of determining based on the image cut method that the present invention proposes among Fig. 5 (c), black part is divided into background, the moving region image sequence I that Fig. 5 (d) shears out
qWidth of cloth in (x, y).Fig. 6 is that three-frame difference and the curve fitting method of automobile video frequency sequence image intensity-based otherness determined the target travel zone.Fig. 6 (a) is the 1st frame of automobile video frequency image, Fig. 6 (b) is the last frame of automobile video frequency image, i.e. the 60th frame, dotted line among Fig. 6 (a), Fig. 6 (b) represents the border of target, the direction of motion that arrow represents target, white portion is the target travel zone of determining based on the image cut method that the present invention proposes among Fig. 6 (c), black part is divided into background, the moving region image sequence I that Fig. 6 (d) shears out
qWidth of cloth in (x, y).Corresponding to the Partial Shear of moving region out, consist of moving region image sequence I in each frame of grayscale image sequence
q(x, y) is with moving region image sequence I
q(x, y) be the input picture of modeling procedure as a setting, and remainder is considered as the background area.
⑸ with moving region image sequence I
qFirst two field picture of (x, y) is as the input data of expectation maximization EM algorithm initialization step.Pedestrian's video motion area image sequence I
qFirst two field picture of (x, y) is one 100 * 145 data space, reading out data point A from this data space
1(1,1) and with A
1Centered by set up rectangular node, the long E of rectangular node
Ps1, wide E
Ps2Computing method are as follows:
Wherein, L is moving region image sequence I
qFirst two field picture of (x, y) shared number of pixels on the x direction of principal axis, H is its number of pixels shared on the y direction of principal axis, k is the mixed Gauss model number of components; For pedestrian's video sequence, L=100, H=145, K=5.
Then rectangular node is with A
1Centered by, length is 20, wide is 29 zone, that is:
Because moving region image sequence I
qIt also can not be decimal that can not there be negative value in first two field picture of (x, y) shared pixel on x axle and y direction of principal axis, is 1 with its assignment when then rectangular node occurs less than 1 value, takes the method that rounds up during decimal.Pedestrian's video image is with A so
1The rectangular node of setting up centered by (1,1) is:
[1，11]×[1，16]
From pedestrian's video motion area image sequence I
qReading data point A successively in first two field picture of (x, y)
2(1,2) judges whether this point falls into A
1Grid in, if can not fall into this grid, then according to the newly-built grid of step said method, the density of grid is 1, if A
2Belong to A
1Grid, corresponding mesh-density increases by 1, repeats above step until pedestrian's video motion area image sequence I
qTill data point in first two field picture of (x, y) is all handled.After finishing above operation, whether judge the density of each grid greater than threshold value, if mesh-density is high fine grid with this grid mark then greater than threshold value den, otherwise with this grid deletion, pedestrian's video sequence den should get 3000.
Appoint and get a high fine grid, judge whether it has the public domain with another high fine grid, think that then these two high fine grids are interconnected if there is such situation, judge again whether newly-generated high fine grid family has the public domain with another high fine grid, if exist then to merge and form new high fine grid family, repeat above step until the high fine grid of all connections consists of high fine grid family, and with the center of gravity of each high fine grid family as initial cluster center.Utilize k average (k-means) algorithm that the input data of EM algorithm initialization step are gathered and be C class, the size of C equals the number of initial cluster center, processes cluster result as the initial parameter value of EM algorithm
∑ and μ, wherein, weights
Equal the ratio of total strong point number of the input data of data point number in each class and EM algorithm initialization step, the variance ∑ equals the covariance of data point in each class, and average μ equals the center of each class.Fig. 7 is that the EM algorithm compares based on the random value initialization with based on grid and the initialized convergence of density estimation method (GEEM) that the present invention proposes, and is as can be seen from the figure faster based on the EM initial method speed of convergence of the present invention's proposition.
⑹ utilize initialized EM algorithm to estimate mixed Gauss model (GMM) parameter, and this process is known technology, may be summarized to be two steps:
(6.1) I. EM algorithm initial value that step ⑸ is obtained, i.e. mixed Gauss model initial parameter value
μ, ∑, substitution formula be the middle moving region image sequence I that calculates 1.
qEach pixel A of (x, y)
iAffiliated gaussian component z
(i)Posterior probability,
Wherein, mixed Gauss model is comprised of a plurality of single Gauss models, and each single Gauss model is called gaussian component, z
(i)The expression pixel A
iAffiliated gaussian component,
The expression pixel A
iThe posterior probability that belongs to j gaussian component;
II. after having estimated posterior probability, utilize formula 2. to recomputate parameter
μ, ∑, and with they be updated to formula 1. in,
Undated parameter
Wherein, m is moving region image sequence I
qThe number of pixel in (x, y), repeating step I, II is until the parameter variation is not remarkable, namely | Θ-Θ ' | (wherein Θ, Θ ' are respectively the parameter before and after upgrading to＜ε
μ, ∑, usually ε=10
-5).
(6.2) for gauss hybrid models, not all gaussian component is all described background.Gauss hybrid models is set up model alike to moving target and background, describes the higher background pixel of the frequency of occurrences with the larger gaussian component of those weight ratios, and the gaussian component of Describing Motion target is then used less weight.According to above principle, with gaussian component according to weights
Order ordering from big to small, the weights addition of b gaussian component is until greater than threshold value before the order, this b gaussian component description is background information so, and what other gaussian component was then described is target information, can fast detecting go out moving target by said method.Be based on as shown in Figure 8 GMM algorithm and based on the result of pedestrian's video sequence image moving object detection of algorithm of the present invention relatively.Wherein Fig. 8 (a), Fig. 8 (b), Fig. 8 (c) are the 1st frame, the 50th frame and the 100th frames of pedestrian's sequence of video images; Fig. 8 (d), Fig. 8 (e), Fig. 8 (f) are that pedestrian's sequence of video images is based on the result of Traditional GM M algorithm moving object detection; Fig. 8 (g), Fig. 8 (h), Fig. 8 (i) be pedestrian's sequence of video images based on the result of algorithm of the present invention (CIGMM) moving object detection, as can be seen from the figure, put forward the methods of the present invention can detect moving target more accurately.
Experiment showed, that the present invention can effectively reduce calculated amount, make algorithm when guaranteeing accuracy, satisfy to greatest extent the requirement of real-time.Experiment condition of the present invention is as follows: allocation of computer is AMD Athlon (tm) Dual Core5200B, 2.69GHz, and the 2GB internal memory, operating system is Microsoft Windows XP, experiment porch is MATLAB7.0.1.Experiment input data be pedestrian's video sequence (120 frames, 320*180) with the vehicle sequence (60 frames, 768*576).Fig. 9 is conventional hybrid Gauss model (GMM), merges the method (EM-GMM) of random initializtion EM and mixed Gauss model and the comparison consuming time that put forward the methods of the present invention (CIGMM) detects moving target.
Claims (2)
1. the moving target detecting method based on improvement mixed Gaussian and image cut is characterized in that, the method may further comprise the steps:
(1) image input step, input raw video image sequence;
(2) image pre-treatment step converts the raw video image sequence to grayscale image sequence;
(3) image difference calculation step:
(3.1) get size and be the moving window W of M * N, W
iThe gray matrix of moving window in the i two field picture of (x, y) expression grayscale image sequence, the gray difference property coefficient R (W of the corresponding window of adjacent two frames in the calculating grayscale image sequence
i(x, y), W
I+1(x, y)), computing method are as follows:
Wherein, x and y are respectively horizontal ordinate and the ordinate of each element in the gray matrix, and M and N are respectively the length of moving window and wide, and 0＜R (W
i(x, y), W
I+1(x, y))＜1;
(3.2) if the gray difference property coefficient greater than threshold value T, then the gray-scale value with moving window correspondence position in the i two field picture of grayscale image sequence all is set as 1, otherwise be set as 0, the span of T is 0.25 to 0.35, mobile moving window is until i frame and the i+1 two field picture of traversal grayscale image sequence, repeat above operation until grayscale image sequence is all finished the computing of the gray scale difference opposite sex, obtain bianry image sequence I
B(x, y);
(3.3) to bianry image sequence I
BAdjacent two frames are ' with ' computing, bianry image sequence I in (x, y)
BAll images obtain target image sequence I after finishing ' with ' computing in (x, y)
M(x, y), I
MGray-scale value is that 1 zone is the preliminary moving target of determining in (x, y);
(4) moving region determining step:
(4.1) according to target image sequence I
MTarget location in (x, y) is determined specifically to comprise the direction of motion of target,
Target image sequence I
M(x, y) comprises the num two field picture altogether, if num is even number, from target image sequence I
MExtract the 1st frame, num/2 two field picture and num two field picture in (x, y), if num is odd number, then from target image sequence I
M(x, y) extract the 1st frame, (num+1)/2 two field picture and num two field picture in, and the peak of target location in 3 two field pictures carried out fitting a straight line, if fitting a straight line and horizontal direction angle are then judged target along continuous straight runs motion less than 45 °, if fitting a straight line and horizontal direction angle judge then that greater than 45 ° target vertically moves;
If target image sequence I
M(x, y) target location in the num two field picture appears at the left side of target location in (num+1)/2 frame or the num/2 frame, the target location appears at the left side of target location in the 1st frame in (num+1)/2 frame or the num/2 frame, judges that then target is to move from right to left; If the target location in the num two field picture appears at the right side of target location in (num+1)/2 frame or the num/2 frame, the target location appears at the right side of target location in the 1st frame in (num+1)/2 frame or the num/2 frame, judges that then target is to move from left to right; Target location in the num two field picture appears at the top of target location in (num+1)/2 frame or the num/2 frame, the target location appears at the top of target location in the 1st frame in (num+1)/2 frame or the num/2 frame, judges that then target is to move from bottom to top; If the target location in the num two field picture appears at the below of target location in (num+1)/2 frame or the num/2 frame, the target location appears at the below of target location in the 1st frame in (num+1)/2 frame or the num/2 frame,, judge that then target is to move from the top down;
(4.2) according to the direction of motion of target, determine object boundary point position, and utilize the method for least square curve fitting to determine the moving region of target;
(4.3) with in each frame of grayscale image sequence corresponding to the Partial Shear of moving region out, consist of moving region image sequence I
q(x, y);
(5) expectation maximization EM algorithm initialization step:
(5.1) moving region image sequence I
qFirst two field picture of (x, y) is as the input data of expectation maximization EM algorithm initialization step, moving region image sequence I
qFirst two field picture of (x, y) is the data space of a L * H, and wherein L is moving region image sequence I
qFirst two field picture of (x, y) shared number of pixels on the x direction of principal axis, H is moving region image sequence I
qFirst two field picture of (x, y) shared number of pixels on the y direction of principal axis;
Read 1 data point A from the input data of EM algorithm initialization step, coordinate is (x, y), and sets up the rectangular node centered by A, and the rectangular node method for building up is as follows:
Rectangular node is centered by A, and long is E
Ps1, wide is E
Ps2The zone, that is:
Wherein,
K is the mixed Gauss model number of components;
(5.2) reading data point successively from the input data of EM algorithm initialization step, if this point can not fall into any one existing grid, then according to the newly-built grid of the method for step (5.1), the density of grid is 1, if the data point of new input belongs to existing grid, corresponding mesh-density increases by 1, until data point is all handled in the input data;
If mesh-density is high fine grid with this grid mark then greater than threshold value den, the value of den is 20% of input data amount check;
(5.3) have the public domain when two high fine grids, perhaps two high fine grids all follow another high fine grid to be communicated with, and think that namely these two high fine grids are interconnected; The highly dense grid cell that is communicated with is combined as high fine grid family, and with the center of gravity of each high fine grid family as initial cluster center;
(5.4) utilize k average (k-means) algorithm that the input data of EM algorithm initialization step are gathered and be C class, the size of C equals the number of initial cluster center in the step (5.3), processes cluster result as the initial parameter value of EM algorithm
∑ and μ, wherein, weights
Equal the ratio of total strong point number of the input data of data point number in each class and EM algorithm initialization step, the variance ∑ equals the covariance of data point in each class, and average μ equals the center of each class;
(6) utilize initialized EM algorithm to estimate mixed Gauss model GMM parameter, then utilize mixed Gauss model to detect moving target, wherein the input data of mixed Gauss model are moving region image sequence I
q(x, y).
2. the method in the middle described definite target travel of step (4.2) zone is according to claim 1:
When the target along continuous straight runs moves, with target image sequence I
M(x, y) peak of target location utilizes least square curve fit to obtain the upper bound in target travel zone in each frame, minimum point utilizes least square curve fit to obtain the lower bound in target travel zone, if target is from right to left motion, the rightest some place vertical line of the first frame, the upper bound lower bound that the most left some place vertical line of last frame and match obtain consists of the moving region of target, if target is from left to right motion, the most left some place vertical line of the first frame, the upper bound lower bound that the rightest some place vertical line of last frame and match obtain consists of the moving region of target;
When target is vertically moved, with target image sequence I
M(x, y) the most left point of target location utilizes least square curve fit to obtain the left margin in target travel zone in each frame, the rightest point utilizes least square curve fit to obtain the right margin in target travel zone, if target is from the top down motion, the first frame peak place horizontal line, the upper bound lower bound that last frame minimum point place horizontal line and match obtain consists of the moving region of target, if target is from bottom to top motion, the first frame minimum point place horizontal line, the upper bound lower bound that last frame peak place horizontal line and match obtain consists of the moving region of target.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2012103659948A CN103077530A (en) | 2012-09-27 | 2012-09-27 | Moving object detection method based on improved mixing gauss and image cutting |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2012103659948A CN103077530A (en) | 2012-09-27 | 2012-09-27 | Moving object detection method based on improved mixing gauss and image cutting |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN103077530A true CN103077530A (en) | 2013-05-01 |
Family
ID=48154049
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2012103659948A Pending CN103077530A (en) | 2012-09-27 | 2012-09-27 | Moving object detection method based on improved mixing gauss and image cutting |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103077530A (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104616298A (en) * | 2015-01-30 | 2015-05-13 | 浙江工业大学之江学院 | Method for detecting moving target of ink-jet printing fabric based on mixed-state Gauss MRF (Markov Random Field) model |
| CN104702378A (en) * | 2013-12-06 | 2015-06-10 | 华为技术有限公司 | Method and device for estimating parameters of mixture Gaussian distribution |
| CN104809098A (en) * | 2014-01-27 | 2015-07-29 | 华为技术有限公司 | Method and device for determining statistical model parameter based on expectation-maximization algorithm |
| CN104820995A (en) * | 2015-04-21 | 2015-08-05 | 重庆大学 | Large public place-oriented people stream density monitoring and early warning method |
| CN105760365A (en) * | 2016-03-14 | 2016-07-13 | 云南大学 | Probability latent parameter estimation model of image semantic data based on Bayesian algorithm |
| CN108038872A (en) * | 2017-12-22 | 2018-05-15 | 中国海洋大学 | One kind perceives follow method based on sound state target detection and Real Time Compression |
| CN108600783A (en) * | 2018-04-23 | 2018-09-28 | 深圳银澎云计算有限公司 | A kind of method of frame rate adjusting, device and terminal device |
| CN108710879A (en) * | 2018-04-20 | 2018-10-26 | 江苏大学 | A kind of pedestrian candidate region generation method based on Grid Clustering Algorithm |
| CN108804995A (en) * | 2018-03-23 | 2018-11-13 | 李春莲 | Device distribution Platform of Image Recognition |
| CN109086647A (en) * | 2018-05-24 | 2018-12-25 | 北京飞搜科技有限公司 | Smog detection method and equipment |
| CN109389073A (en) * | 2018-09-29 | 2019-02-26 | 北京工业大学 | The method and device of detection pedestrian area is determined by vehicle-mounted camera |
| CN110348305A (en) * | 2019-06-06 | 2019-10-18 | 西北大学 | A kind of Extracting of Moving Object based on monitor video |
| CN111583357A (en) * | 2020-05-20 | 2020-08-25 | 重庆工程学院 | Object motion image capturing and synthesizing method based on MATLAB system |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20050147277A1 (en) * | 2004-01-05 | 2005-07-07 | Honda Motor Co., Ltd | Apparatus, method and program for moving object detection |
| CN101098465A (en) * | 2007-07-20 | 2008-01-02 | 哈尔滨工程大学 | Moving object detecting and tracing method in video monitor |
-
2012
- 2012-09-27 CN CN2012103659948A patent/CN103077530A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20050147277A1 (en) * | 2004-01-05 | 2005-07-07 | Honda Motor Co., Ltd | Apparatus, method and program for moving object detection |
| CN101098465A (en) * | 2007-07-20 | 2008-01-02 | 哈尔滨工程大学 | Moving object detecting and tracing method in video monitor |
Non-Patent Citations (1)
| Title |
|---|
| JINFU YANG ET AL.: "An Efficient Moving Object Detection Algorithm based on Improved GMM and Cropped Frame Technique", 《PORCEEDINGS OF 2012 IEEE INTERNATIONAL CONFERENCE ON MECHATRONICS AND AUTOMATION》 * |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104702378A (en) * | 2013-12-06 | 2015-06-10 | 华为技术有限公司 | Method and device for estimating parameters of mixture Gaussian distribution |
| CN104702378B (en) * | 2013-12-06 | 2018-03-09 | 华为技术有限公司 | The method for parameter estimation and device of Gaussian mixtures |
| CN104809098A (en) * | 2014-01-27 | 2015-07-29 | 华为技术有限公司 | Method and device for determining statistical model parameter based on expectation-maximization algorithm |
| WO2015109781A1 (en) * | 2014-01-27 | 2015-07-30 | 华为技术有限公司 | Method and device for determining parameter of statistical model on the basis of expectation maximization |
| CN104616298B (en) * | 2015-01-30 | 2017-04-19 | 浙江工业大学之江学院 | Method for detecting moving target of ink-jet printing fabric based on mixed-state Gauss MRF (Markov Random Field) model |
| CN104616298A (en) * | 2015-01-30 | 2015-05-13 | 浙江工业大学之江学院 | Method for detecting moving target of ink-jet printing fabric based on mixed-state Gauss MRF (Markov Random Field) model |
| CN104820995A (en) * | 2015-04-21 | 2015-08-05 | 重庆大学 | Large public place-oriented people stream density monitoring and early warning method |
| CN105760365A (en) * | 2016-03-14 | 2016-07-13 | 云南大学 | Probability latent parameter estimation model of image semantic data based on Bayesian algorithm |
| CN108038872A (en) * | 2017-12-22 | 2018-05-15 | 中国海洋大学 | One kind perceives follow method based on sound state target detection and Real Time Compression |
| CN108038872B (en) * | 2017-12-22 | 2021-08-31 | 中国海洋大学 | Dynamic and static target detection and real-time compressed sensing tracking research method |
| CN108804995B (en) * | 2018-03-23 | 2019-04-16 | 新昌县夙凡软件科技有限公司 | Toilet facility distributed image identifying platform |
| CN108804995A (en) * | 2018-03-23 | 2018-11-13 | 李春莲 | Device distribution Platform of Image Recognition |
| CN108710879A (en) * | 2018-04-20 | 2018-10-26 | 江苏大学 | A kind of pedestrian candidate region generation method based on Grid Clustering Algorithm |
| CN108600783B (en) * | 2018-04-23 | 2021-03-30 | 深圳齐心好视通云计算有限公司 | Frame rate adjusting method and device and terminal equipment |
| CN108600783A (en) * | 2018-04-23 | 2018-09-28 | 深圳银澎云计算有限公司 | A kind of method of frame rate adjusting, device and terminal device |
| CN109086647A (en) * | 2018-05-24 | 2018-12-25 | 北京飞搜科技有限公司 | Smog detection method and equipment |
| CN109086647B (en) * | 2018-05-24 | 2022-01-07 | 苏州飞搜科技有限公司 | Smoke detection method and device |
| CN109389073A (en) * | 2018-09-29 | 2019-02-26 | 北京工业大学 | The method and device of detection pedestrian area is determined by vehicle-mounted camera |
| CN110348305A (en) * | 2019-06-06 | 2019-10-18 | 西北大学 | A kind of Extracting of Moving Object based on monitor video |
| CN110348305B (en) * | 2019-06-06 | 2021-06-25 | 西北大学 | Moving object extraction method based on monitoring video |
| CN111583357A (en) * | 2020-05-20 | 2020-08-25 | 重庆工程学院 | Object motion image capturing and synthesizing method based on MATLAB system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103077530A (en) | Moving object detection method based on improved mixing gauss and image cutting | |
| CN106407889B (en) | Method for recognizing human body interaction in video based on optical flow graph deep learning model | |
| CN101470809B (en) | Moving object detection method based on expansion mixed gauss model | |
| CN104680559B (en) | The indoor pedestrian tracting method of various visual angles based on motor behavior pattern | |
| CN101729872B (en) | Video monitoring image based method for automatically distinguishing traffic states of roads | |
| CN103106667B (en) | A kind of towards blocking the Moving Objects method for tracing with scene change | |
| CN103530893B (en) | Based on the foreground detection method of background subtraction and movable information under camera shake scene | |
| CN104200485A (en) | Video-monitoring-oriented human body tracking method | |
| CN103258332B (en) | A kind of detection method of the moving target of resisting illumination variation | |
| CN111814621A (en) | Multi-scale vehicle and pedestrian detection method and device based on attention mechanism | |
| CN104820997B (en) | A kind of method for tracking target based on piecemeal sparse expression Yu HSV Feature Fusion | |
| CN105528794A (en) | Moving object detection method based on Gaussian mixture model and superpixel segmentation | |
| CN111027505B (en) | Hierarchical multi-target tracking method based on significance detection | |
| CN103488993A (en) | Crowd abnormal behavior identification method based on FAST | |
| CN101408983A (en) | Multi-object tracking method based on particle filtering and movable contour model | |
| CN109919053A (en) | A kind of deep learning vehicle parking detection method based on monitor video | |
| CN103020614B (en) | Based on the human motion identification method that space-time interest points detects | |
| CN104537684A (en) | Real-time moving object extraction method in static scene | |
| CN106023245A (en) | Static background moving object detection method based on neutrosophy set similarity measurement | |
| CN104599291B (en) | Infrared motion target detection method based on structural similarity and significance analysis | |
| CN106780560A (en) | A kind of feature based merges the bionic machine fish visual tracking method of particle filter | |
| CN102063727A (en) | Covariance matching-based active contour tracking method | |
| CN115830265A (en) | Automatic driving movement obstacle segmentation method based on laser radar | |
| CN103456029A (en) | Mean Shift tracking method for resisting similar color and illumination variation interference | |
| CN106056078A (en) | Crowd density estimation method based on multi-feature regression ensemble learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C02 | Deemed withdrawal of patent application after publication (patent law 2001) | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20130501 |