CN107038710B - It is a kind of using paper as the Vision Tracking of target - Google Patents
It is a kind of using paper as the Vision Tracking of target Download PDFInfo
- Publication number
- CN107038710B CN107038710B CN201710082445.2A CN201710082445A CN107038710B CN 107038710 B CN107038710 B CN 107038710B CN 201710082445 A CN201710082445 A CN 201710082445A CN 107038710 B CN107038710 B CN 107038710B
- Authority
- CN
- China
- Prior art keywords
- sample
- characteristic value
- sampling
- block
- frame
- 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.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of using paper as the Vision Tracking of target, the classification of positive negative sample is carried out in first frame image, calculate the characteristic value of all positive samples and negative sample, and determine a target area, every frame image after the second frame is sampled around the target area of previous frame with sampling block, just in conjunction with previous frame, the characteristic value of negative sample calculates the score of all sampling blocks, obtain a point highest sampling block, LSD line detection algorithm is utilized to each sampling block, find that there are [70 ° of angle, 110 °] and two straight-line intersections inside the sampling block, further select two straight-line intersections near a paracentral sampling block as object block, considering paper in this way has shooting angle with camera in moving process, so that finally obtained tracking target is more accurate.
Description
Technical field
The present invention relates to computer vision tracking technique fields, and in particular to a kind of tracked using paper as the vision of target is calculated
Method.
Background technique
Application of the target following of view-based access control model in fields such as video monitoring, human-computer interactions is gradually popularized, and this respect is ground
Study carefully be also computer vision field research hotspot.Most application scenarios of current this visual target tracking all have non-
Normal grain details abundant, such as face, pedestrian and vehicle, these grain details provide enough features for target following
Information constitutes the feature calculation space of target following, although since illumination or posture etc. cause feature jump to will affect tracking
Robustness.But for some particularly simple scenes, such as paper or card target in desk tops, because of texture information pole
It is deficient, and how to track equally is the project for being worth further investigation.
Existing track algorithm depends critically upon the characteristic point detected, and for this patent, characteristic point quantity is seriously not
Foot causes the robustness of tracking inadequate.This algorithm obtains the tracking of entire paper, robust by strengthening the tracking of paper apex angle
Property is greatly enhanced, and robust tracking still may be implemented in the case where being partially blocked.
Summary of the invention
In order to solve the deficiencies in the prior art, the present invention provides one kind, is achieved by the following technical programs:
It is a kind of to choose the rectangular paper of a texture-free information using paper as the Vision Tracking of target, recording paper
This paper is placed on the desktop of texture-free information by size, which is arbitrarily moved and rotated on the table, in real time simultaneously
Continuously acquire the video image that paper carries out random movement and rotary course on the table.
Step 1: obtaining the first frame of video image, an angle of the paper is chosen, by the center of mark window and the angle
Vertex be overlapped, and by the mark window it is each in each with paper it is parallel, obtain a marked region, record the mark zone
The location information of domain central point and using the rectangular area as target area;Choose the mark window to the region where the angle into
Any label of row obtains multiple sample images, obtains m positive sample and n negative samples according to the determination method of positive sample and negative sample
This;
The determination method of positive sample and the negative sample are as follows: calculate the central point of sample image and the center of target area
The distance of point, the sample image by distance less than h are labeled as positive sample, and the sample image apart from h is greater than and less than v is labeled as
Negative sample image, the h take [4,8], and v takes [20,25];
Step 2: the feature of the positive sample and negative sample in first frame image is calculated separately using the calculation method of characteristic value
Value;
The feature value calculating method are as follows:
For any one positive sample or negative sample, the current frame image where the sample is converted to grayscale image, to this
Grayscale image carries out gradient calculating, obtains the gradient-norm of the gray level image, constructs the integrogram of the gradient-norm;
It is utilized respectively t feature templates to sample the sample, obtains t sampling area, calculated using integrogram
The pixel of each sampling area gradient-norm and it isThe feature of each positive sample and negative sample is calculated separately using formula (1)
Value, wherein the value range of t is 50-200;
Wherein, f indicates characteristic value in formula (1),K=0.06, IxIt is the frame image in x-axis side
Upward change rate, IyThe change rate of frame image in the y-axis direction;
Step 3: calculating expectation of the characteristic value in i-th dimension degree of all positive samples in video image first frameAnd standard
DifferenceCalculate expectation of the characteristic value in i-th dimension degree of all negative samplesAnd standard deviationWherein i value is 1-50;
Step 4: obtaining the second frame of video image as present frame, according to sampling block create-rule on current frame image
It generates m sampling block and calculates each sampling block using feature value calculating method using each sampling block as a positive sample
Characteristic value, wherein m takes [500,1000];The score that each sampling block is calculated using score computation method obtains score most
Ten high sampling blocks;
The sampling block create-rule are as follows:
For current frame image, multiple sampling blocks are generated at random around the target area that previous frame determines, described adopts
Sample block size mark window size is identical, and the target area that the sampling block and previous frame determine is empty in the pixel of the frame image
Between upper there is overlapping;
The score computation method are as follows:
The value H (v) that each sampling block is calculated using formula (4) is score;
Wherein n indicates the number of dimensions of characteristic value, n=50;
Wherein p (vi| y=1) be the i-th dimension degree that the sampling block is positive sample conditional probability, pass through formula (2) calculate;Wherein, fiFor characteristic value of the sampling block under i-th dimension degree,With
For all positive sample characteristic values in previous frame video figure i-th dimension degree expectation and standard deviation;
p(vi| y=0) it be the conditional probability of the i-th dimension degree that the sample is negative sample is p (vi| y=0), pass through formula (3)
It calculates;Wherein fiFor characteristic value of the sampling block under i-th dimension degree,WithFor all negative sample characteristic values in previous frame video image i-th dimension degree expectation and standard deviation;
Step 5: method, which finds needs, to be determined using object block for ten sampling blocks of the highest scoring that step 5 obtains
One sampling block is as object block;Each sampling block will be obtained in step 4 as a sample image, the object block that will be obtained
As a target area, tracking is distinguished the positive sample in current frame image according to positive sample and the determination method of negative sample and is born
Sample updates the expectation and standard deviation of present frame all positive samples and negative sample according to expectation and standard deviation update method, obtains
Updated positive sample and negative sample are in the expectation and standard deviation of i-th dimension degree, wherein i value arbitrary integer between 1-50;
The determination method of the object block are as follows:
For each of ten sampling blocks of highest scoring, it is utilized respectively LSD straight-line detection, obtains each sampling block
Straight line existing for inside finds inside there are two included angle of straight line at [70 °, 110 °] and two straight-line intersections are in the sampling block
Internal sampling block alternately object block, selects two straight-line intersections near a paracentral alternative target block as target
Block;
The expectation and standard deviation update method are as follows:
The characteristic value of positive sample and negative sample is obtained according to feature value calculating method, calculates separately positive sample and negative sample
Expectation and standard deviation of the characteristic value in i-th dimension degree;Existed according to the characteristic value that formula (5) calculate the positive sample of updated present frame
The expectation and standard deviation of i-th dimension, wherein i value is 1-50;
In formula (5)Indicate expectation of the characteristic value in i-th dimension of previous frame positive sample,Be present frame just
The characteristic value of sample i-th dimension expectation,It is expectation of the characteristic value in i-th dimension of updated present frame positive sample;It is standard deviation of the characteristic value in i-th dimension of previous frame positive sample,It is that the characteristic value of present frame positive sample exists
The standard deviation of i-th dimension,It is standard deviation of the characteristic value in i-th dimension of present frame positive sample after updating, wherein 0 < λ < 1;
Expectation and standard deviation of the characteristic value in i-th dimension of the negative sample of present frame are updated according to formula (6);
In formula (6)Indicate expectation of the characteristic value in i-th dimension of previous frame negative sample,It is negative for present frame
The characteristic value of sample i-th dimension expectation,It is expectation of the characteristic value in i-th dimension of updated present frame negative sample,Indicate standard deviation of the characteristic value in i-th dimension of previous frame negative sample,It is the characteristic value of present frame negative sample
In the standard deviation of i-th dimension,It is standard deviation of the characteristic value in i-th dimension of updated present frame negative sample.
Step 6: for its excess-three angle of paper, repeating step 1-5, find the corresponding object block in each angle of paper, root
According to the center of the corresponding object block of four apex angles of paper in the position in current frame image, four apex angle positions of paper are obtained
It sets, determines the position of paper in this frame image;
Step 7: obtaining next each frame of video image as present frame, repeat step 4,5, obtain the frame image
In four object blocks, obtain four corner positions of position of the center of object block in current frame image as paper, root
The position of paper in this frame image is determined according to four corner positions;
When wherein repeating step (4), when generating sampling block, the object block that previous frame is determined is as sampling block production method
In target area;
Wherein when repeating step 4, in formula (2)WithFor updated all positive samples in previous frame video figure
Characteristic value i-th dimension degree expectation and standard deviation, in formula (3)WithIt is updated all in previous frame video image
Expectation and standard deviation of the characteristic value of negative sample in i-th dimension degree.
Wherein, the size of the mark window is 32*32, and the m takes arbitrary integer between 4000-6000, and n takes
Arbitrary integer between 1000-2000.
Wherein, the feature template is the random side length that rectangle and length and width are respectively less than mark window, each feature template
A random factor W is generated at random.
Wherein, the step 5, in 6,7, if utilizing LSD in ten sampling blocks of the corresponding highest scoring of some apex angle
The straight line inside each sampling block that straight-line detection is found, for this ten sampling blocks, not finding inside, there are two straight lines
Angle [70 °, 110 °] sampling block, though find two included angle of straight line this two straight lines in [70 °, 110 °] range
When intersection point is not inside the sampling block, then it is assumed that do not find object block, at this time according to parallelogram law, utilize other 3
The object block of apex angle determines the object block at this angle.
It has following technical effect that above technical scheme is compared with the prior art
1, in this algorithm, the sampling block of ten highest scorings is chosen in sampling block, it is straight using LSD to each sampling block
Line detection algorithms find there are angle [70 °, 110 °] and two straight-line intersections inside the sampling block, further select two
Straight-line intersection near a paracentral sampling block as object block, consider in this way paper in moving process with camera
There are problems that shooting angle, so that finally obtained tracking target is more accurate.
2, in this algorithm, when calculating sampling block eigenvalue, it is to be carried out in the gradient modular space of gray level image, eliminates very
More interference informations enhance the calculating of characteristic value using angle point response when calculating characteristic value, such as formula (1), so that diagonal
Tracking is strengthened, the final robustness for improving the tracking of paper apex angle.
3, in this algorithm, the tracking for entire paper is realized by the tracking of 4 apex angles to paper.If with
Track process, some apex angle tracking failure, i.e., do not find the corresponding object block of the apex angle, then directly utilize parallelogram method
Then, the position of the apex angle is directly calculated using the center of the corresponding object block of other three apex angles.It may finally improve
The robustness that paper is tracked in the case where being at least partially obscured.
Detailed description of the invention
Fig. 1 is the flow chart of this algorithm;
Fig. 2 is target area schematic diagram of this algorithm in the first frame video image acceptance of the bid note;
Fig. 3 is schematic diagram of this algorithm in multiple sample images of the first frame video image acceptance of the bid note;
Fig. 4 is 8 kinds of result schematic diagrams that this algorithm is likely to be obtained sampling block using LSD straight-line detection;
The sampling block for ten highest scorings that Fig. 5, which is this algorithm, to be obtained in third frame image carries out the knot of LSD straight-line detection
Fruit;
Fig. 6 is that the method determined using object block finds straight line signal inside the corresponding object block in certain angle in third frame image
Figure;
Fig. 7 is the schematic diagram of the object block at four angles of paper obtained using this algorithm;
Fig. 8 is the schematic diagram that another object block is determined according to the corresponding object block in three angles of paper.
Specific embodiment
In the present invention in the location information of record tracking target, for each frame video image, with a left side for video image
Upper angle is coordinate origin, using horizontal direction as Y-axis, establishes coordinate system by X-axis of vertical direction, record tracking target is in the coordinate
Coordinate in system, using the coordinate as the location information of tracking target.
The present invention chooses the rectangular paper of one texture-free information, this paper is placed on the desktop of texture-free information, will
The paper is arbitrarily moved and is rotated on the table, in real time and is continuously acquired paper and is arbitrarily moved and rotated on the table
The video image of process;
The first frame of video image is obtained, the size of the mark window of selection is 32*32, if Fig. 2 is first frame video figure
The tracking object delineation marked as in, h take 8, v to take 20, obtain 4506 positive samples and 1364 negative samples, generate 50 at random
A feature template.
In score computation method of the invention, in calculating the second frame image when the score of sampling block, The expectation and standard deviation of the positive sample and negative sample that are determined respectively in first frame, from third frame and later
When every frame calculates sampling block score, useIt is that present frame is updated, it is expected that can with standard deviation
It is calculated by data.
Embodiment 1
In the present embodiment, the rectangle paper for choosing a blank is placed on a white smooth surface (texture-free information),
The paper is arbitrarily moved and rotated on the table, in real time and paper is continuously acquired and is arbitrarily moved and revolved on the table
Turn over the video image of journey.
The first frame for obtaining video image chooses an angle of the paper, and the size of the mark window of selection is 32*32,
The center of mark window is overlapped with the vertex at the angle, and by the mark window it is each in each with paper it is parallel, obtain one
A marked region records the location information of the marked region central point and using the rectangular area as target area;If Fig. 2 is the
The target area schematic diagram marked in one frame video image, when determining positive sample and negative sample, h takes 8, v to take 20, as Fig. 3 exists
The schematic diagram of multiple sample images of first frame video image acceptance of the bid note determines sample further according to the determination method of positive negative sample
Positive sample and negative sample in image obtain 5234 positive samples according to the determination method of positive sample and negative sample and 1456 negative
Sample.
The characteristic value for calculating positive sample and negative sample in first frame image, in the present embodiment when calculating sample image, elder generation
Current frame image is converted to grayscale image, gradient calculating is carried out to the grayscale image, obtains the gradient-norm of the gray level image, further
The integrogram of the gradient-norm is constructed, and constructs 120 feature templates and each positive sample or negative sample is sampled,
According to integrogram calculate each sample image pixel and, reduce in this way algorithm traversal number, improve computational efficiency.
Simultaneously when calculating sampling block eigenvalue, it is to be carried out in the gradient modular space of gray level image, eliminates many interference
Information enhances the calculating of characteristic value using angle point response when calculating characteristic value, such as formula (1), so that diagonal tracks
To reinforcing, the final robustness for improving the tracking of paper apex angle.
Calculate separately all positive samples and the characteristic value of negative sample in video image first frame under 1-50 dimension every
The expectation and standard deviation of dimension;
Second frame for obtaining video image is used as present frame, the pixel space position in the target area of first frame determination
Surrounding generates 800 sampling blocks at random, and the size of the sampling block of generation and the size of mark window are identical, i.e. the size of sampling block
With positive sample, the same size of negative sample, the location information of target area is obtained in first frame, generates sampling block in present frame
When, it is necessary to assure there is overlapping, the sampling generated in this way in pixel space in the target area in the sampling block of generation and upper frame
Block is exactly potential positive sample;While in order to further find object block in these sampling blocks;Pass through calculating in the present embodiment
The score of all sampling blocks selects the sampling block of ten highest scorings.
A kind of determination method for tracking target is provided in the present embodiment, calculates the sampling block of ten highest scorings, it is right
What it is in the highest scoring that third frame obtains is that the score of sampling block is respectively, 1227,1216,1210,1203,1199,1188,
1183,1179,1168,1136, LSD straight-line detection is carried out to each sampling block;Because even one of highest scoring samples
Block and is not centainly the object block that we need yet, because carrying out each sampling block by straight-line detection, obtaining result may
There are eight kinds of situations, if Fig. 4 is to carry out 8 kinds of result schematic diagrams that straight-line detection is likely to be obtained to sampling block using LSD in this algorithm,
As seen from the figure, the form of straight lines detected inside all sampling blocks is different, shares eight kinds of situations.
In this algorithm, straight-line detection is carried out to ten sampling blocks of the highest scoring that third frame image obtains, as a result such as Fig. 5
It is shown, it can be seen that relatively, in Fig. 5, first sampling block two straight for form of straight lines existing for inside this ten sampling blocks
Wire clamp angle is 82 °, and the angle of straight line, close to the center of sampling block, the angle of the sampling block of the 2nd, 3,5,6,8,9 and 10 also exists
In [70 °, 110 °] range, but straight-line intersection and first sampling block are compared to the center for deviateing sampling block, the 7th sampling block angle
It is 125 °, not in the range of we require.
If find at this time it is internal there are two included angle of straight line [70 °, 110 °] and two straight-line intersections in the sampling block
Inside considers paper in this way and exists in moving process with camera and shoot then by the sampling block alternately object block
The problem of angle, when finding to take [70 °, 110 °] after many experiments, obtains so the corner angle of paper may change
Result it is more accurate so that finally obtained tracking target is more accurate;Selected in alternative target two straight-line intersections near
Paracentral one is used as object block;The object block that is found using the method is as shown in fig. 6, straight inside sampling block in Fig. 6
Wire clamp angle is 82 °, and is compared with other nine sampling blocks, and the intersection point of two inside the sampling block straight line is in sampling block
The heart, then using the sampling block as object block.
Do not find it is internal there are two included angle of straight line at [70 °, 110 °] though or find two straight lines, two straight lines
When intersection point is not inside the sampling block, then it is assumed that do not find object block.The sampling block an of highest scoring is taken to make with existing
It is compared for tracking target, provided in this embodiment to be detected using ten sampling blocks of the LSD straight-line detection to highest scoring, energy
More accurate determines tracking target, while considering in paper movement and rotary course, the shooting angle of camera and paper
Degree problem, two included angle of straight line obtained by straight-line detection are more reasonable at [70 °, 110 °], while in order to enable arriving
It is more accurate to track target, selected in alternative features block in the present embodiment two straight-line intersections near sampling block center as
Track target.
The prior art only obtains a point highest characteristic block just as tracking target, and the tracking target obtained in this way is accurate
Property it is not high because by straight-line detection, the form of the straight line inside ten characteristic blocks of highest scoring, is different, thus this
Invention is repeated by the way that two included angle of straight line of acquisition are at [70 °, 110 °] and intersection point is near the conduct object block at sampling block center
The above method finds the corresponding object block in each angle of paper, if Fig. 7 is paper four in certain the frame image found according to the present embodiment
The corresponding object block in a angle can determine according to using the center of object block as the position at the paper angle by four object blocks
The position of paper in this frame image.
In the present embodiment, in the position for determining paper according to the center of the corresponding object block of several apex angles found
It, can be according to the mesh at other angles having determined if there is when not finding corresponding object block of one or two angles when information
Block is marked, in conjunction with parallelogram law and known paper size, to determine the corresponding object block in other two angle, and then is determined
Location information of the paper in current frame video image, the object block at such obtained each angle are accurately, to improve
The robustness of algorithm, according to the location information of these three tracking targets, is utilized if Fig. 8 has looked for the corresponding tracking target in three angles
Parallelogram law can determine the location information at the last one angle of paper in conjunction with known paper size.
Every frame video image can find satisfactory characteristic block as tracking target mostly, and only a few cases can go out
Corresponding tracking target is not found in some existing angle.
Each frame image after the second frame, after object block determines, in order to which the determining object block of next frame is more accurate,
It needs to be updated the expectation of the positive negative sample of the frame and standard deviation, the expectation and standard deviation in conjunction with previous frame are to phase of this frame
It hopes and standard deviation is reconciled, in this way when next frame calculates the score of each sampling block, the expectation of the previous frame of use and mark
Quasi- difference is exactly, so that the object block of score is more accurate, to improve the accuracy of track algorithm after being updated.
Claims (4)
1. a kind of using paper as the Vision Tracking of target, choose the rectangular paper of a texture-free information, recording paper it is big
It is small, this paper is placed on the desktop of texture-free information, which is arbitrarily moved and rotated on the table, in real time and even
The continuous video image for obtaining paper and carrying out random movement and rotary course on the table, which is characterized in that
Step 1: obtaining the first frame of video image, an angle of the paper is chosen, by the top at the center of mark window and the angle
Point is overlapped, and by the mark window it is each in each with paper it is parallel, obtain a marked region, record in the marked region
The location information of heart point and using the marked region as target area;The mark window is chosen to appoint the region where the angle
Meaning label obtains multiple sample images, obtains m positive sample and n negative sample according to the determination method of positive sample and negative sample;
The determination method of positive sample and the negative sample are as follows: calculate the central point of sample image and the central point of target area
The sample image that distance is less than h is labeled as positive sample by distance, and the sample image by distance greater than h and less than v marks the sample that is negative
This image, the h take [4,8], and v takes [20,25];
Step 2: the characteristic value of the positive sample and negative sample in first frame image is calculated separately using the calculation method of characteristic value;
The feature value calculating method are as follows:
For any one positive sample or negative sample, the current frame image where the sample is converted to grayscale image, to the gray scale
Figure carries out gradient calculating, obtains the gradient-norm of the grayscale image, constructs the integrogram of the gradient-norm;
It is utilized respectively t feature templates to sample the sample, obtains t sampling area, calculated using integrogram each
The pixel of sampling area gradient-norm and it isThe characteristic value of each positive sample and negative sample is calculated separately using formula (1),
Wherein, the value range of t is 50-200;
Wherein, f indicates characteristic value in formula (1),K=0.06, IxIn the direction of the x axis for the frame image
Change rate, IyThe change rate of frame image in the y-axis direction;
Step 3: calculating expectation of the characteristic value in i-th dimension degree of all positive samples in video image first frameAnd standard deviation
Calculate expectation of the characteristic value in i-th dimension degree of all negative samplesAnd standard deviationWherein i value is 1-50;
Step 4: obtaining the second frame of video image as present frame, generated on current frame image according to sampling block create-rule
M sampling block calculates the spy of each sampling block using feature value calculating method using each sampling block as a positive sample
Value indicative, wherein m takes [500,1000];The score that each sampling block is calculated using score computation method, obtains highest scoring
Ten sampling blocks;
The sampling block create-rule are as follows:
For current frame image, multiple sampling blocks, the sampling block are generated at random around the target area that previous frame determines
Size mark window size is identical, and the target area that the sampling block and previous frame determine is in the pixel space of the frame image
There are overlappings;
The score computation method are as follows:
The value H (v) that each sampling block is calculated using formula (4) is score;
Wherein n indicates the number of dimensions of characteristic value, n=50;
Wherein p (vi| y=1) be the i-th dimension degree that the sampling block is positive sample conditional probability, pass through formula (2) calculate;Wherein, fiFor characteristic value of the sampling block under i-th dimension degree,With
For all positive sample characteristic values in previous frame video figure i-th dimension degree expectation and standard deviation;
p(vi| y=0) be the i-th dimension degree that the sample is negative sample conditional probability, pass through formula (3) calculate;Wherein fiFor characteristic value of the sampling block under i-th dimension degree,With
For all negative sample characteristic values in previous frame video image i-th dimension degree expectation and standard deviation;
Step 5: method, which finds one of needs, to be determined using object block for ten sampling blocks of the highest scoring that step 5 obtains
Sampling block is as object block;Each sampling block will be obtained in step 4 as a sample image, by obtained object block as
Positive sample and negative sample in current frame image are distinguished in one target area, tracking according to positive sample and the determination method of negative sample
This, the expectation and standard deviation of present frame all positive samples and negative sample are updated according to expectation and standard deviation update method, is obtained more
Positive sample and negative sample after new are in the expectation and standard deviation of i-th dimension degree, wherein i value arbitrary integer between 1-50;
The determination method of the object block are as follows:
For each of ten sampling blocks of highest scoring, it is utilized respectively LSD straight-line detection, is obtained inside each sampling block
Existing straight line finds inside there are two included angle of straight line at [70 °, 110 °] and two straight-line intersections are inside the sampling block
Sampling block alternately object block, select two straight-line intersections near a paracentral alternative target block as object block;
The expectation and standard deviation update method are as follows:
The characteristic value of positive sample and negative sample is obtained according to feature value calculating method, calculates separately the feature of positive sample and negative sample
It is worth the expectation and standard deviation in i-th dimension degree;The characteristic value of the positive sample of updated present frame is calculated i-th according to formula (5)
The expectation and standard deviation of dimension, wherein i value is 1-50;
In formula (5)Indicate expectation of the characteristic value in i-th dimension of previous frame positive sample,It is present frame positive sample
Characteristic value i-th dimension expectation,It is expectation of the characteristic value in i-th dimension of updated present frame positive sample;It is standard deviation of the characteristic value in i-th dimension of previous frame positive sample,It is that the characteristic value of present frame positive sample exists
The standard deviation of i-th dimension,It is standard deviation of the characteristic value in i-th dimension of present frame positive sample after updating, wherein 0 < λ < 1;
Expectation and standard deviation of the characteristic value in i-th dimension of the negative sample of present frame are updated according to formula (6);
In formula (6)Indicate expectation of the characteristic value in i-th dimension of previous frame negative sample,For present frame negative sample
Characteristic value i-th dimension expectation,It is expectation of the characteristic value in i-th dimension of updated present frame negative sample,Indicate standard deviation of the characteristic value in i-th dimension of previous frame negative sample,It is the characteristic value of present frame negative sample
In the standard deviation of i-th dimension,It is standard deviation of the characteristic value in i-th dimension of updated present frame negative sample;
Step 6: for its excess-three angle of paper, repeating step 1-5, the corresponding object block in each angle of paper is found, according to paper
The center of the corresponding object block of four apex angles is opened in the position in current frame image, obtains four corner positions of paper, really
Determine the position of paper in this frame image;
Step 7: obtaining next each frame of video image as present frame, repeat step 4,5, obtain in the frame image
Four object blocks obtain four corner positions of position of the center of object block in current frame image as paper, according to four
A corner position determines the position of paper in this frame image;
When wherein repeating step 4, when generating sampling block, the object block that previous frame is determined is as the mesh in sampling block production method
Mark region;
Wherein when repeating step 4, in formula (2)WithFor all positive sample features updated in previous frame video figure
It is worth the expectation and standard deviation in i-th dimension degree, in formula (3)WithFor all negative samples updated in previous frame video image
Expectation and standard deviation of this characteristic value in i-th dimension degree.
2. as described in claim 1 using paper as the Vision Tracking of target, which is characterized in that in the step 1,
The size of the mark window is 32*32, and m takes arbitrary integer between 4000-6000, and n takes any whole between 1000-2000
Number.
3. as described in claim 1 using paper as the Vision Tracking of target, which is characterized in that the feature template is square
Shape and length and width are respectively less than the random side length of mark window, each feature template generates a random factor W at random.
4. as described in claim 1 using paper as the Vision Tracking of target, which is characterized in that the step 5,6,7
In, if in ten sampling blocks of the corresponding highest scoring of some apex angle, inside each sampling block for being found using LSD straight-line detection
Straight line, for this ten sampling blocks, do not find it is internal there are two included angle of straight line [70 °, 110 °] sampling block, or
Though finding two included angle of straight line when this interior two straight-line intersections of [70 °, 110 °] range are not inside the sampling block, then it is assumed that
Object block is not found, and at this time according to parallelogram law, the target at this angle is determined using the object block of other 3 apex angles
Block.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710082445.2A CN107038710B (en) | 2017-02-15 | 2017-02-15 | It is a kind of using paper as the Vision Tracking of target |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710082445.2A CN107038710B (en) | 2017-02-15 | 2017-02-15 | It is a kind of using paper as the Vision Tracking of target |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107038710A CN107038710A (en) | 2017-08-11 |
CN107038710B true CN107038710B (en) | 2019-07-02 |
Family
ID=59533521
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710082445.2A Expired - Fee Related CN107038710B (en) | 2017-02-15 | 2017-02-15 | It is a kind of using paper as the Vision Tracking of target |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107038710B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109708658B (en) * | 2019-01-14 | 2020-11-24 | 浙江大学 | Visual odometer method based on convolutional neural network |
CN110097561B (en) * | 2019-03-14 | 2022-07-15 | 长安大学 | Rapid paper detection and segmentation method based on space constraint conditions |
CN113781525B (en) * | 2021-09-13 | 2024-01-23 | 陕西铁路工程职业技术学院 | Three-dimensional target tracking method based on original CAD model |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102750772A (en) * | 2012-06-05 | 2012-10-24 | 广东智华计算机科技有限公司 | Paper money tracking system and method based on machine vision |
CN106127203A (en) * | 2016-06-29 | 2016-11-16 | 孟祥雨 | A kind of device to object location and followed the trail of and the method for image recognition |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20070247422A1 (en) * | 2006-03-30 | 2007-10-25 | Xuuk, Inc. | Interaction techniques for flexible displays |
-
2017
- 2017-02-15 CN CN201710082445.2A patent/CN107038710B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102750772A (en) * | 2012-06-05 | 2012-10-24 | 广东智华计算机科技有限公司 | Paper money tracking system and method based on machine vision |
CN106127203A (en) * | 2016-06-29 | 2016-11-16 | 孟祥雨 | A kind of device to object location and followed the trail of and the method for image recognition |
Non-Patent Citations (2)
Title |
---|
"Real-Time Compressive Tracking";Kaihua Zhang etal.;《IEEE Computer Society Press》;20121231;第866-877页 |
"动态变形表面的四维测量";史金龙 等;《计算机工程与应用》;20111231;第202-206页 |
Also Published As
Publication number | Publication date |
---|---|
CN107038710A (en) | 2017-08-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10607362B2 (en) | Remote determination of containers in geographical region | |
CN108564616B (en) | Fast robust RGB-D indoor three-dimensional scene reconstruction method | |
CN109829398B (en) | Target detection method in video based on three-dimensional convolution network | |
CN106683091B (en) | A kind of target classification and attitude detecting method based on depth convolutional neural networks | |
CN104850850B (en) | A kind of binocular stereo vision image characteristic extracting method of combination shape and color | |
CN108171752A (en) | A kind of sea ship video detection and tracking based on deep learning | |
CN112766184B (en) | Remote sensing target detection method based on multi-level feature selection convolutional neural network | |
CN104156965B (en) | A kind of automatic quick joining method of Mine Monitoring image | |
CN103886325B (en) | Cyclic matrix video tracking method with partition | |
CN105427298A (en) | Remote sensing image registration method based on anisotropic gradient dimension space | |
CN109035300B (en) | Target tracking method based on depth feature and average peak correlation energy | |
CN107038710B (en) | It is a kind of using paper as the Vision Tracking of target | |
CN111145228A (en) | Heterogeneous image registration method based on local contour point and shape feature fusion | |
CN109101981A (en) | Winding detection method based on global image bar code under a kind of streetscape scene | |
CN106530313A (en) | Sea-sky line real-time detection method based on region segmentation | |
EP3553700A2 (en) | Remote determination of containers in geographical region | |
CN102446356A (en) | Parallel and adaptive matching method for acquiring remote sensing images with homogeneously-distributed matched points | |
CN113468968A (en) | Remote sensing image rotating target detection method based on non-anchor frame | |
CN108960267A (en) | System and method for model adjustment | |
CN106709870A (en) | Close-range image straight-line segment matching method | |
CN116310837A (en) | SAR ship target rotation detection method and system | |
Shuai et al. | An improved YOLOv5-based method for multi-species tea shoot detection and picking point location in complex backgrounds | |
CN113011359A (en) | Method for simultaneously detecting plane structure and generating plane description based on image and application | |
Yaosheng et al. | Object tracking in satellite videos based on improved correlation filters | |
CN110414379A (en) | In conjunction with the building extraction algorithm of elevation map Gabor textural characteristics and LiDAR point cloud feature |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
TA01 | Transfer of patent application right | ||
TA01 | Transfer of patent application right |
Effective date of registration: 20171221 Address after: 710064 Shaanxi Province, Xi'an Yanta District No. 126 Middle Ring Road Applicant after: Changan Univ. Applicant after: SHAANXI RAILWAY INSTITUTE Address before: 710064 Shaanxi Province, Xi'an Yanta District No. 126 Middle Ring Road Applicant before: Changan Univ. |
|
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190702 Termination date: 20200215 |