CN106650668A - Method and system for detecting movable target object in real time - Google Patents
Method and system for detecting movable target object in real time Download PDFInfo
- Publication number
- CN106650668A CN106650668A CN201611223198.5A CN201611223198A CN106650668A CN 106650668 A CN106650668 A CN 106650668A CN 201611223198 A CN201611223198 A CN 201611223198A CN 106650668 A CN106650668 A CN 106650668A
- Authority
- CN
- China
- Prior art keywords
- image
- real
- subwindow
- scanning
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/20—Scenes; Scene-specific elements in augmented reality scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/285—Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a method and system for detecting a movable target object in real time, and the method comprises the following steps: collecting and making samples according to the photographed images with a target object, carrying out the classification of the samples, and obtaining a classifier; obtaining an image of a target region in real time, and obtaining a to-be-detected image; carrying out the continuous sliding of a scanning child window in the to-be-detected image or a to-be-detected image region, wherein the scanning child window calculates the MB-LBP feature of a region when arriving at a position, carrying out the classification of the features according to the class, and determining a target region which meets the detection conditions of the target object of the classifier, so as to obtain the target object. The method achieves a purpose of detecting the movable target object in real time, and can provide coordinates, which can be moved and combined with the reality, for an AR product.
Description
Technical field
The present invention with regard to machine learning techniques field, the method for more particularly to a kind of real-time detection movable object thing and
System.
Background technology
Augmented reality (Augmented Reality, abbreviation AR), refers to and be superimposed on the basis of real scene virtual information
And it is displayed on the terminals out to realize the seamless connection of real world and virtual world, form man-machine interaction, make the two generation
Boundary is perfectly combined together in user's eye.The interesting and interactive of juvenile product for adding AR technologies is all greatly improved,
The feeling of immersion of child can be strengthened, be conducive to cultivating the concentration of child, be beneficial to cultivate child to modern science, outfield
Interest, be conducive to the exploitation of intelligence, it is deep to be welcome by masses.
The virtual world and real world of AR performances with reference to when need " interface " of real world, " interface " here
It is exactly the coordinate of certain in esse object in reality.It is with reference to the virtual scene for setting up AR with this coordinate.Current AR
It is typically immovable as the object of coordinate object of reference in order to ensure stablizing for virtual scene in product, even if some AR
The coordinate object of reference of product can be moved, also since it is considered that the identification with reference to object is positioned, the scene or reference to application
The lightness colors of object have certain restriction, make the design and use of product and have significant limitation.
The content of the invention
To overcome the shortcomings of that above-mentioned prior art is present, the purpose of the present invention is to provide a kind of real-time detection removable mesh
The method and system of mark thing, using realize by movable object thing as AR products coordinate object of reference, and suitable for indoor and outdoor
Several scenes under various illumination.
It is that, up to above and other purpose, the present invention proposes a kind of method of real-time detection movable object thing, including as follows
Step：
Step one, according to the image collection with target object for shooting sample is made, and sample is classified, and is divided
Class device；
Step 2, obtains in real time the image of target area, obtains altimetric image to be checked；
Step 3, is constantly slided using one scan subwindow in image or image-region to be detected to be detected, should
The every position calculation of scanning subwindow goes out the MB-LBP features in the region, and this feature is classified according to the classification, sentences
Break the target area for meet grader to object testing conditions, obtain object.
Further, after step 3, also comprise the steps：
If obtaining multiple objects in step 3, optimal solution is selected from multiple objects using canny operators.
Further, the step of this selects optimal solution using canny operators from multiple objects further includes：
The initial screening of object is carried out to multiple objects using canny operators characteristic value；
According to object color set in advance, the maximum object of corresponding color area ratio is selected as optimal solution.
Further, the maximum object of white area ratio is selected as optimal solution.
Further, step 3 is further included：
Step S1, judges whether previous frame detects object；
Step S2, if previous frame detects object, according to the testing result of previous frame present frame object is determined
Size and the position being likely to occur, determine the scanning subwindow size and detection range of present frame, using the scanning subwindow in
Slide to obtain object in detection range；
Step S3, it is continuous in current frame image region using scanning subwindow if previous frame is not detected by object
Slide to obtain object.
Further, in step S2, using between frame and frame relevance prediction present frame object size and
The position being likely to occur.
To reach above-mentioned purpose, the present invention also provides a kind of system of real-time detection movable object thing, including：
Object grader sets up unit, right for making sample according to the image collection with target object for shooting
Sample is classified, and obtains grader；
Image acquisition unit, for the image for obtaining target area in real time, obtains altimetric image to be checked；
Target following detector unit, using one scan subwindow in image or image-region to be detected to be detected not
Disconnected to slide, the every position calculation of the scanning subwindow goes out the MB-LBP features in the region, this feature is entered according to the classification
Row classification, determines the target area for meeting grader to object testing conditions；
Further, the system also include optimal objective thing screening unit, for using canny operators from multiple objects
In select optimal solution.
Further, the optimal objective thing screening unit obtains as follows optimal solution：
The initial screening of object is carried out to multiple objects using canny operators characteristic value；
According to object color set in advance, the maximum object of corresponding color area ratio is selected as optimal solution.
Further, whether the target following detector unit first determines whether previous frame for each frame altimetric image to be checked
Object is detected, if previous frame detects object, the big of present frame object is determined according to the testing result of previous frame
Position that is little and being likely to occur, determines the size and detection range of the scanning subwindow of present frame, and using the scanning subwindow
Slide to obtain object in detection range, if previous frame is not detected by object, using scanning subwindow in present frame
Constantly slide to detect object in image-region.
Compared with prior art, a kind of method and system of real-time detection movable object thing of the invention are by gathering training
Sample, obtains object grader, using the target area in image of the object grader to real-time detection determine whether for
Object, and optimal objective thing is filtered out from multiple objects, the purpose of real-time detection movable object thing is realized, can be
AR products provide the coordinate for movably combining with reality, and the present invention utilizes machine learning techniques, can be by the object for recognizing
Positional information be supplied to upper layer software (applications) do AR scenes building process, the present invention based on machine learning object identification prevent
Wind robustness is high, it is adaptable to the several scenes under the various illumination in indoor and outdoor, compares similar with irremovable and restriction background
AR products have more flexibility.
Description of the drawings
The step of Fig. 1 is a kind of method of real-time detection movable object thing of the invention flow chart；
Fig. 2 is the thin portion flow chart of step 101 in the specific embodiment of the invention；
Fig. 3 is the schematic diagram for choosing target object minimum rectangle part in the specific embodiment of the invention in step S11；
Fig. 4 is the schematic diagram of a sample needed for the specific embodiment of the invention；
Fig. 5 is the image schematic diagram that the origin that mobile terminal is obtained in the specific embodiment of the invention is the lower left corner；
Fig. 6 is the image schematic diagram after in the specific embodiment of the invention Fig. 5 images is spun upside down；
Fig. 7 is to detect the schematic diagram that multiple regions meet grader in the specific embodiment of the invention；
Fig. 8 is canny operator edge detection result schematic diagrams in the specific embodiment of the invention；
Fig. 9 is the mean value of the canny operator edge detection figures of the target area detected in the specific embodiment of the invention
Schematic diagram；
Figure 10 is target area schematic diagram after preliminary screening in the specific embodiment of the invention；
Figure 11 is that optimal solution schematic diagram is obtained in the specific embodiment of the invention；
The successive frame of the video image that Figure 12-Figure 14 is provided for mobile terminal in the specific embodiment of the invention；
Figure 15 is the schematic diagram that the specific embodiment of the invention determines present frame target location according to previous frame testing result；
Figure 16 is the detection process flow chart in the specific embodiment of the invention to object；
Figure 17 is a kind of system architecture diagram of the system of real-time detection movable object thing of the invention；
Figure 18 is the detail structure chart that object grader sets up unit 170 in the specific embodiment of the invention.
Specific embodiment
Below by way of specific instantiation and embodiments of the present invention are described with reference to the drawings, those skilled in the art can
The further advantage and effect of the present invention are understood easily by content disclosed in the present specification.The present invention also can be different by other
Instantiation implemented or applied, the every details in this specification also can based on different viewpoints with application, without departing substantially from
Various modifications and change are carried out under the spirit of the present invention.
The step of Fig. 1 is a kind of method of real-time detection movable object thing of the invention flow chart.As shown in figure 1, this
A kind of method of bright real-time detection movable object thing, comprises the steps：
Step 101, according to the image collection with target object for shooting sample is made, and sample is classified, and is obtained
The grader of training sample.
Fig. 2 is the thin portion flow chart of step 101 in the specific embodiment of the invention.As shown in Fig. 2 step 101 is specifically included：
Step S11, according to the image collection with target object for shooting sample is made.In the specific embodiment of the invention
In, step S11 is further included：
Step1, shoots the image with object, and choosing in the captured image with target object to frame
The minimum rectangle part of whole target object, as shown in Figure 3.
Step2, by the image of selection the image of n × n-pixel is converted to, and the image after conversion is to be obtained one
Sample.In the specific embodiment of the invention, the image (i.e. the image of the rectangle size in Fig. 3) chosen is converted to into 30 × 30 pictures
The image of element, as shown in figure 4, the image shown in Fig. 4 is a resulting sample.
Step3, repeats Step 1, Step 2, until required sample number is obtained, generally with the sample of more than 2000
Training can just obtain comparatively ideal grader file.
Step S12, to sample feature extraction is carried out, and chooses the feature for training sample.
Because MB-LBP features (Muti-Block Local Binary Pattern) have gray scale consistency and rotation not
The advantage of denaturation, and training speed will be many soon than other features (such as Haar and HOG), therefore the present invention selects MB-LBP
Feature of the feature as training sample.
MB-LBP operators are the big squares to 3 × 3, by the average gray g for comparing wherein center blockmWith the center sub-block
Average gray { the g of surrounding eight neighborhood sub-block1, g2..., g8Calculate characteristic value.Specific formula for calculation is as follows:
Wherein
Step S13, using the feature extracted sample training is carried out, and sample is classified, and obtains the grader of object.
In the present invention, choose either mode sorting algorithm (including and be not limited to such as decision tree, Bayesian network, people
Artificial neural networks, K- neighbours, SVMs, Adaboost algorithm, Boosting algorithms etc.) sample is classified.
In the specific embodiment of the invention, choose Adaboost (Adaptive Boosting) algorithm and illustrate sample
Classification declaration, Adaboost algorithm can strengthen the weight of the sample of previous basic classification device misclassification, the whole after weighting
Sample is used to train next basic classification device again, meanwhile, a new Weak Classifier, Zhi Daoda are added in each wheel
To certain predetermined sufficiently small error rate or reach preassigned maximum iteration time and terminate.
It is assumed that training dataset is T={ (x1, y1), (x2, y2) ... (xn, yn), wherein xiIt is sample instance, yiFor mark
Note, positive sample is labeled as+1, and negative sample is labeled as -1, i=1,2 ..., n.Wherein positive sample is following Step 1, in Step 2
The sample of the target sectional drawing of making, negative sample is the arbitrary figure not comprising target.Concrete classifying step is as follows：
Step 1, initializes the weights of training sample, and each training sample most starts to be endowed identical weight：
Specifically,
D1(i)=1/n (formula 3)
Step 2, training is using according to D1The Weak Classifier C of (i) sampling1.Here Weak Classifier C1Input for training
Sample, output+1 when grader classification results are positive sample, grader classification results are output as -1 when being negative sample.
Step 3, calculates C1Error.
E1=∑ D1(i)I[C1(xi)≠yi] (formula 4)
Step 4, according to C1Error calculation C1Corresponding weight
Step 5, updates training sample weight
Step 6, calculates the sample weights of kth time iteration, the error of Weak Classifier Ck, and then obtains Weak Classifier CkIt is right
The weight answered.
Wherein, the sample weights of kth time iteration are：
Weak Classifier CkError be：
Ek=∑ Dk(i)I[Ck(xi)≠yi] (formula 8)
Weak Classifier CkCorresponding weight is：
Wherein k=1,2 ..., m.
Step8, according to Weak Classifier CkAnd its weight, obtain grader.In the specific embodiment of the invention, finally give
The function of classification samples x：
The classification results of sample are：
Step 102, obtains in real time the image of target area, obtains altimetric image to be checked.
In the specific embodiment of the invention, the image of target area is obtained in real time using the post-positioned pick-up head of mobile device.
Because the origin of the system default image of mobile terminal is the lower left corner, and in general pattern identification facility (such as opencv)
The origin of coordinates of image is the upper left corner, so the image seen at opencv ends is the appearance of the turned upside down shown in Fig. 5, identification
It is front to need the image of Fig. 5 to be spun upside down into shown in Fig. 6.
Step 103, is constantly slided using a scanning subwindow in image or image-region to be detected to be detected
Dynamic, the every position of the scanning subwindow then calculates the MB-LBP features in the region, is then trained according to step 101
Grader g (x) is classified to this feature, judges whether the region meets testing conditions of the grader to object.
In the specific embodiment of the invention, the length-width ratio for scanning subwindow is consistent with the length-width ratio of training sample, correspondingly from
The minimum window size of setting to each size of maximum window size will create corresponding subwindow, treat detection image
Region is traveled through, therefore is controlled scanning subwindow size range and judged that the region that target is likely to occur can reduce scanning in advance
Number of times, improves detection speed.
Step 104, from multiple objects optimal solution is selected.
As shown in fig. 7, sometimes piece image has multiple regions and meets testing conditions of the grader to target, and this is applied
In known only one of which target, so need find out the target object really to be detected in the multiple regions for detecting.
Specifically, step 104 further includes following steps：
Step 1, using canny operators characteristic value the initial screening of object is carried out.
In the specific embodiment of the invention, because the object to be detected is that the overwhelming majority is white Similar Round Object,
Rectangular area shared by target do average obtained from the rim detection of canny operators can in a fixed region, according to
Experimental result before, it is recognised that generally this numerical value is in the interval of [30,60].As in Fig. 8, j to r is respectively a to i
Canny operator edge detection results.
The numeral shown in Fig. 9 is respectively the mean value of the canny operator edge detection figures of the target area for detecting.Its
In, it is the region excluded in preliminary screening that average is 26.2,60.7,63.7,64.9,85.1 region.
Step 2, selects the maximum object of white area ratio as optimal solution.
After the screening of Step 1, tetra- regions of a also shown in Figure 10, c, e, g.According in HSV color spaces
Three variables do binaryzation to candidate region.
V=max (formula 14)
Wherein, (r, g, b) is corresponding three values of the corresponding rgb color space of coloured image some coordinate, max
R, g are equivalent to respectively with min, the maximum and minimum of a value in b.(h, s, v) is three that corresponding HSV space represents color
Value.R in above formula, g, b ∈ [0,255], and h ∈ [0,360), h, s ∈ [0,1].
In opencv, H ∈ scopes [0,179], H, S ∈ [0,255], we S values in [0,30], V values [200,
255] Partial filtration between is out, as a result b, d, f, the white portion in h.The ratio in the region wherein shared by white portion
It is respectively 0%, 25.3%, 50.9%, 0%.In sum, optimal solution is region e.The area of the spheroid solid box i.e. shown in Figure 11
Domain e.
It is preferred that step 103 further includes following steps：
Step S31, judges whether previous frame detects object；
Step S32, if previous frame detects object, according to the testing result of previous frame scanning of present frame is determined
Window size and detection range, and enter line slip detection in detection range using the scanning subwindow, to obtain object；
Step S33, it is continuous in current frame image region using scanning subwindow if previous frame is not detected by object
Slide, to obtain object.
Specifically, the scanning subwindow often scans to position calculation the MB-LBP features for going out the region, according to this
Classification is classified to this feature, judges to meet target area of the grader to object testing conditions, to obtain object.
In the specific embodiment of the invention, because application scenarios are real-time scenes, below figure 12-14 show mobile terminal
The image of offer is the successive frame of video, there is very high similarity between frame and frame, therefore can be from the testing result of previous frame
The size for determining present frame target and the position being likely to occur, can so avoid the scanning in the range of whole image, reduce
Detection time, realizes target following.
By taking Figure 15 as an example, it is assumed that the position for detecting target in present frame is a width of w shown in solid box, the square of a height of h
Shape region.In view of the relevance between frame and frame, while target object and handheld terminal can be moved rapidly, can be next frame
Detection range be located at centered on the center of present frame, it is a height of to represent present frame detects the high h in rectangular area of target three
Times, equally, three times of a width of this rectangle region field width w, dashed rectangle region as shown in figure 15.
So, the present invention can be with using the position that is likely to occur of relevance prediction next frame target object between frame and frame
The time greatly saved required for detection, improve verification and measurement ratio.For example, the size of scanning subwindow of the invention is change
's.The minimum subwindow size of such as setting is 20*20, is 200*200 to the maximum, and the multiple that scanning window changes every time is 2, then
Scanning subwindow size is 20*20 when scanning to first window of a certain testing image, and the upper left angle sweep from image is to image
The lower right corner, if testing image size 480*640, this time scanning times are (480-20+1) * (640-20+1)=286281
It is secondary；Window size is changed into 40*40 when second window is scanned, this time scanning times be (480-40+1) * (640-40+1)=
265041 times；Third time window size is 80*80 ..., and last time window size is 200*200, scanning times
For (480-200+1) * (640-200+1)=123921, so total scanning times are 286281+265041+ ...+123921 times；
And if the target sizes that previous frame is detected are 80*80, and knowing position, then the scanning to present frame can be minimum chi
Very little to be set to 40*40, full-size is set to 160*160, and detection range is set to (80*3) * (80*3), then only need to use 40*40,80*
80,1,60*,160 tri- scanning windows.During 40*40 windows, scanning times are (240-40+1) * (240-40+1)=40401；80*
During 80 window, scanning times are (240-80+1) * (240-80+1)=25921；During 160*160 windows, scanning times are (240-
160+1) * (240-160+1)=6561, so total scanning times 40401+25921+6561 is then much smaller than 286281+
265041+ ...+123921, greatlys save the time required for detection.
Figure 16 is the detection process flow chart in the specific embodiment of the invention to object.First, initialize, loaded targets
Grader, arranges relevant parameter, such as including the full-size of detection target, minimum dimension, and each change of scanning window times
Number, grader document location；Judge whether previous frame detects object；If having detected that, from the resulting estimate of previous frame
Go out the position range that target in present frame is likely to occur, and the size of scanning subwindow and the model of scanning are determined according to estimation result
Enclose；Constantly slided in image or image-region to be detected to be detected using scanning subwindow, the scanning subwindow is every
The MB-LBP features in the region are then calculated to position, then this feature are classified according to the grader for training,
Judge whether the region meets testing conditions of the grader to object, to detect object；Whether the object for detecting
For multiple；If multiple, then screening meets the target of textured condition, selects the most target of white area, and is back to target
Position, until frame of video reads finishing.
Figure 17 is a kind of system architecture diagram of the system of real-time detection movable object thing of the invention.As shown in figure 17, originally
A kind of system of real-time detection movable object thing is invented, including：Object grader sets up unit 170, image acquisition unit
171st, target following detector unit 172 and optimal objective thing screening unit 173.
Object grader sets up unit 170, for making sample according to the image collection with target object for shooting,
Sample is classified, the grader of training sample is obtained.
Figure 18 is the detail structure chart that object grader sets up unit 170 in the specific embodiment of the invention.Such as Figure 18 institutes
Show, object grader is set up unit 170 and specifically included：
Sample collection unit 1701, for making sample according to the image collection with target object for shooting.At this
In bright specific embodiment, sample collection unit 1701 collects as follows sample：
Step1, shoots the image with object, and choosing in the captured image with target object to frame
The minimum rectangle part of whole target object.
Step2, by the image of selection the image of n × n-pixel is converted to, and the image after conversion is to be obtained one
Sample.In the specific embodiment of the invention, the image of selection is converted to into the image of 30 × 30 pixels, that is, obtains required one
Sample.
Step3, repeats Step 1, Step 2, until required sample number is obtained, generally with the sample of more than 2000
Training can just obtain comparatively ideal grader file.
Feature extraction unit 1702, for carrying out feature extraction to sample, chooses the feature for training sample.
Because MB-LBP features (Muti-Block Local Binary Pattern) have gray scale consistency and rotation not
The advantage of denaturation, and training speed will be many soon than other features (such as Haar and HOG), therefore the present invention selects MB-LBP
Feature of the feature as training sample.
MB-LBP operators are the big squares to 3 × 3, by the average gray g of Correlation Centre sub-blockmWith eight neighborhood around it
Average gray { the g of sub-block1, g2..., g8Calculate characteristic value.Specific formula for calculation is as follows:
Wherein
Taxon 1703, using the feature extracted sample training is carried out, and sample is classified, and obtains grader.
In the present invention, choose either mode sorting algorithm (including and be not limited to such as decision tree, Bayesian network, people
Artificial neural networks, K- neighbours, SVMs, Adaboost algorithm, Boosting algorithms etc.) sample is classified.
In the specific embodiment of the invention, choose Adaboost (Adaptive Boosting) algorithm and illustrate sample
Classification declaration, Adaboost algorithm can strengthen the weight of the sample of previous basic classification device misclassification, the whole after weighting
Sample is used to train next basic classification device again, meanwhile, a new Weak Classifier, Zhi Daoda are added in each wheel
To certain predetermined sufficiently small error rate or reach preassigned maximum iteration time and terminate.
Image acquisition unit 171, for the image for obtaining target area in real time, obtains altimetric image to be checked.
In the specific embodiment of the invention, the image of target area is obtained in real time using the post-positioned pick-up head of mobile device.
Because the origin of the system default image of mobile terminal is the lower left corner, and in general pattern identification facility (such as opencv)
The origin of coordinates of image is the upper left corner, therefore needs to spin upside down into image before recognizing.
Target following detector unit 172, using a scanning subwindow in image or image district to be detected to be detected
Constantly slide in domain, the every position of the scanning subwindow then calculates the MB-LBP features in the region, then according to mesh
Mark thing grader sets up the grader that unit 170 trains and this feature is classified, and judges whether the region meets grader
Testing conditions to object.
In the specific embodiment of the invention, the length-width ratio for scanning subwindow is consistent with the length-width ratio of training sample, correspondingly from
The minimum window size of setting to each size of maximum window size will create corresponding subwindow, treat detection image
Region is traveled through, therefore is controlled scanning subwindow size range and judged that the region that target is likely to occur can reduce scanning in advance
Number of times, improves detection speed.
In the specific embodiment of the invention, because application scenarios are real-time scenes, there is very high similarity between frame and frame,
Therefore can the size for determining present frame target from the testing result of previous frame and the position being likely to occur, so can avoid
Scanning in the range of whole image, reduces detection time, realizes target following.Therefore target following detector unit 172 is for every
One two field picture, can judge whether previous frame detects object first, if previous frame detects object, according to previous frame
Size that testing result determines present frame object and the position being likely to occur, and determine present frame scanning subwindow size and
Detection range, is constantly slided using scanning subwindow in current frame image region, the every position meter of the scanning subwindow
The MB-LBP features in the region are calculated, this feature is classified according to the classification, judge to meet grader to target quality testing
The target area of survey condition.The position that the present invention is likely to occur using the relevance prediction next frame target object between frame and frame
The time that can greatly save required for detection, improve verification and measurement ratio.
Optimal objective thing screening unit 173, for selecting in multiple objects for detecting from target following detector unit 172
Go out optimal solution.
Sometimes piece image has multiple regions and meets testing conditions of the grader to target, and in this application it is known only
One target, so needing to find out the target object really to be detected in the multiple regions for detecting.
Specifically, optimal objective thing screening unit 173 carries out as follows optimal objective thing screening：
Step 1, using canny operators characteristic value the initial screening of object is carried out.
In the specific embodiment of the invention, because the object to be detected is that the overwhelming majority is white Similar Round Object,
Rectangular area shared by target do average obtained from the rim detection of canny operators can in a fixed region, according to
Experimental result before, it is recognised that generally this numerical value is in the interval of [30,60].
Step 2, selects the maximum object of white area ratio as optimal solution.
In sum, a kind of method and system of real-time detection movable object thing of the invention pass through to gather training sample,
Object grader is obtained, is determined whether for target using the target area in image of the object grader to real-time detection
Thing, and optimal objective thing is filtered out from multiple objects, the purpose of real-time detection movable object thing being realized, can produce for AR
Product provide the coordinate for movably combining with reality, and the present invention utilizes machine learning techniques, can be by the position of the object for recognizing
Confidence breath is supplied to the building that upper layer software (applications) does AR scenes to process, identification windproof Shandong of the present invention based on the object of machine learning
Rod is high, it is adaptable to the several scenes under the various illumination in indoor and outdoor, compares and is produced with similar AR that is irremovable and limiting background
Product have more flexibility.
The principle and its effect of above-described embodiment only illustrative present invention, it is of the invention not for limiting.Any
Art personnel can be modified above-described embodiment and are changed under the spirit and the scope without prejudice to the present invention.Therefore,
The scope of the present invention, should be as listed by claims.
Claims (10)
1. a kind of method of real-time detection movable object thing, comprises the steps：
Step one, according to the image collection with target object for shooting sample is made, and sample is classified, and is classified
Device；
Step 2, obtains in real time the image of target area, obtains altimetric image to be checked；
Step 3, is constantly slided using one scan subwindow in image or image-region to be detected to be detected, the scanning
The every position calculation of subwindow goes out the MB-LBP features in the region, and this feature is classified according to the classification, judges
Meet target area of the grader to object testing conditions, obtain object.
2. a kind of method of real-time detection movable object thing as claimed in claim 1, it is characterised in that in step 3 it
Afterwards, also comprise the steps：
If obtaining multiple objects in step 3, optimal solution is selected from multiple objects using canny operators.
3. a kind of method of real-time detection movable object thing as claimed in claim 2, it is characterised in that this utilizes canny
The step of operator selects optimal solution from multiple objects further includes：
The initial screening of object is carried out to multiple objects using canny operators characteristic value；
According to object color set in advance, the maximum object of corresponding color area ratio is selected as optimal solution.
4. a kind of method of real-time detection movable object thing as claimed in claim 3, it is characterised in that select white area
The maximum object of ratio is used as optimal solution.
5. a kind of method of real-time detection movable object thing as claimed in claim 1, it is characterised in that step 3 is further
Including：
Step S1, judges whether previous frame detects object；
Step S2, if previous frame detects object, according to the testing result of previous frame the size of present frame object is determined
And the position being likely to occur, the scanning subwindow size and detection range of present frame are determined, using the scanning subwindow in detection
In the range of slide to obtain object；
Step S3, if previous frame is not detected by object, is constantly slided using scanning subwindow in current frame image region
To obtain object.
6. a kind of method of real-time detection movable object thing as claimed in claim 5, it is characterised in that：In step S2,
The size that the object of present frame is predicted using the relevance between frame and frame and the position being likely to occur.
7. a kind of system of real-time detection movable object thing, including：
Object grader sets up unit, for making sample according to the image collection with target object for shooting, to sample
Classified, obtained grader；
Image acquisition unit, for the image for obtaining target area in real time, obtains altimetric image to be checked；
Target following detector unit, is constantly slided using one scan subwindow in image or image-region to be detected to be detected
Dynamic, the every position calculation of the scanning subwindow goes out the MB-LBP features in the region, this feature is carried out point according to the classification
Class, determines the target area for meeting grader to object testing conditions.
8. a kind of system of real-time detection movable object thing as claimed in claim 7, it is characterised in that：The system also includes
Optimal objective thing screening unit, for selecting optimal solution from multiple objects using canny operators.
9. a kind of system of real-time detection movable object thing as claimed in claim 8, it is characterised in that the optimal objective
Thing screening unit obtains as follows optimal solution：
The initial screening of object is carried out to multiple objects using canny operators characteristic value；
According to object color set in advance, the maximum object of corresponding color area ratio is selected as optimal solution.
10. a kind of system of real-time detection movable object thing as claimed in claim 7, it is characterised in that：The target with
Track detector unit first determines whether whether previous frame detects object for each frame altimetric image to be checked, if previous frame is detected
Object, the then size for present frame object being determined according to the testing result of previous frame and the position being likely to occur, it is determined that currently
The size and detection range of the scanning subwindow of frame, and slide to obtain target in detection range using the scanning subwindow
Thing, if previous frame is not detected by object, constantly slides to detect using scanning subwindow in current frame image region
Object.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611223198.5A CN106650668A (en) | 2016-12-27 | 2016-12-27 | Method and system for detecting movable target object in real time |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611223198.5A CN106650668A (en) | 2016-12-27 | 2016-12-27 | Method and system for detecting movable target object in real time |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106650668A true CN106650668A (en) | 2017-05-10 |
Family
ID=58831423
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611223198.5A Pending CN106650668A (en) | 2016-12-27 | 2016-12-27 | Method and system for detecting movable target object in real time |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106650668A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107995442A (en) * | 2017-12-21 | 2018-05-04 | 北京奇虎科技有限公司 | Processing method, device and the computing device of video data |
| CN108171157A (en) * | 2017-12-27 | 2018-06-15 | 南昌大学 | The human eye detection algorithm being combined based on multiple dimensioned localized mass LBP histogram features with Co-HOG features |
| CN108664032A (en) * | 2018-06-08 | 2018-10-16 | 东北大学秦皇岛分校 | A kind of automatic control system and method for following carrier based on machine vision |
| CN109685002A (en) * | 2018-12-21 | 2019-04-26 | 创新奇智(广州)科技有限公司 | A kind of dataset acquisition method, system and electronic device |
| CN110009000A (en) * | 2019-03-11 | 2019-07-12 | 东北大学 | The grain heap object detection method of sorting algorithm is improved based on ADABOOST+SVM |
| CN110032947A (en) * | 2019-03-22 | 2019-07-19 | 深兰科技（上海）有限公司 | A kind of method and device that monitor event occurs |
| CN111382613A (en) * | 2018-12-28 | 2020-07-07 | 中国移动通信集团辽宁有限公司 | Image processing method, apparatus, device and medium |
| CN111739055A (en) * | 2020-06-10 | 2020-10-02 | 新疆大学 | Infrared point-like target tracking method |
| CN112184756A (en) * | 2020-09-30 | 2021-01-05 | 北京理工大学 | Single-target rapid detection method based on deep learning |
| CN112639684A (en) * | 2018-09-11 | 2021-04-09 | 苹果公司 | Method, device and system for delivering recommendations |
| CN112926722A (en) * | 2021-01-27 | 2021-06-08 | 上海兰宝传感科技股份有限公司 | Method for counting people in escalator entrance area |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102004925A (en) * | 2010-11-09 | 2011-04-06 | 无锡中星微电子有限公司 | Method for training object classification model and identification method using object classification model |
| CN103661102A (en) * | 2012-08-31 | 2014-03-26 | 北京旅行者科技有限公司 | Method and device for reminding passersby around vehicles in real time |
| US20140169663A1 (en) * | 2012-12-19 | 2014-06-19 | Futurewei Technologies, Inc. | System and Method for Video Detection and Tracking |
| CN104937638A (en) * | 2013-01-22 | 2015-09-23 | 高通股份有限公司 | Systems and methods for tracking and detecting a target object |
| CN105787888A (en) * | 2014-12-23 | 2016-07-20 | 联芯科技有限公司 | Human face image beautifying method |
-
2016
- 2016-12-27 CN CN201611223198.5A patent/CN106650668A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102004925A (en) * | 2010-11-09 | 2011-04-06 | 无锡中星微电子有限公司 | Method for training object classification model and identification method using object classification model |
| CN103661102A (en) * | 2012-08-31 | 2014-03-26 | 北京旅行者科技有限公司 | Method and device for reminding passersby around vehicles in real time |
| US20140169663A1 (en) * | 2012-12-19 | 2014-06-19 | Futurewei Technologies, Inc. | System and Method for Video Detection and Tracking |
| CN104937638A (en) * | 2013-01-22 | 2015-09-23 | 高通股份有限公司 | Systems and methods for tracking and detecting a target object |
| CN105787888A (en) * | 2014-12-23 | 2016-07-20 | 联芯科技有限公司 | Human face image beautifying method |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107995442A (en) * | 2017-12-21 | 2018-05-04 | 北京奇虎科技有限公司 | Processing method, device and the computing device of video data |
| CN108171157A (en) * | 2017-12-27 | 2018-06-15 | 南昌大学 | The human eye detection algorithm being combined based on multiple dimensioned localized mass LBP histogram features with Co-HOG features |
| CN108664032A (en) * | 2018-06-08 | 2018-10-16 | 东北大学秦皇岛分校 | A kind of automatic control system and method for following carrier based on machine vision |
| CN112639684A (en) * | 2018-09-11 | 2021-04-09 | 苹果公司 | Method, device and system for delivering recommendations |
| CN109685002A (en) * | 2018-12-21 | 2019-04-26 | 创新奇智(广州)科技有限公司 | A kind of dataset acquisition method, system and electronic device |
| CN111382613A (en) * | 2018-12-28 | 2020-07-07 | 中国移动通信集团辽宁有限公司 | Image processing method, apparatus, device and medium |
| CN111382613B (en) * | 2018-12-28 | 2024-05-07 | 中国移动通信集团辽宁有限公司 | Image processing method, device, equipment and medium |
| CN110009000A (en) * | 2019-03-11 | 2019-07-12 | 东北大学 | The grain heap object detection method of sorting algorithm is improved based on ADABOOST+SVM |
| CN110009000B (en) * | 2019-03-11 | 2022-09-02 | 东北大学 | Grain pile target detection method based on ADABOOST + SVM improved classification algorithm |
| CN110032947B (en) * | 2019-03-22 | 2021-11-19 | 深兰科技（上海）有限公司 | Method and device for monitoring occurrence of event |
| CN110032947A (en) * | 2019-03-22 | 2019-07-19 | 深兰科技（上海）有限公司 | A kind of method and device that monitor event occurs |
| CN111739055B (en) * | 2020-06-10 | 2022-07-05 | 新疆大学 | Infrared point-like target tracking method |
| CN111739055A (en) * | 2020-06-10 | 2020-10-02 | 新疆大学 | Infrared point-like target tracking method |
| CN112184756A (en) * | 2020-09-30 | 2021-01-05 | 北京理工大学 | Single-target rapid detection method based on deep learning |
| CN112926722A (en) * | 2021-01-27 | 2021-06-08 | 上海兰宝传感科技股份有限公司 | Method for counting people in escalator entrance area |
| CN112926722B (en) * | 2021-01-27 | 2022-11-01 | 上海兰宝传感科技股份有限公司 | Method for counting people in escalator entrance area |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106650668A (en) | Method and system for detecting movable target object in real time | |
| CN109147254B (en) | Video field fire smoke real-time detection method based on convolutional neural network | |
| CN102708361B (en) | Human face collecting method at a distance | |
| CN100393106C (en) | Method and apparatus for detecting and/or tracking image or color area of image sequence | |
| CN105335725B (en) | A kind of Gait Recognition identity identifying method based on Fusion Features | |
| Graf et al. | Multi-modal system for locating heads and faces | |
| CN106997597B (en) | It is a kind of based on have supervision conspicuousness detection method for tracking target | |
| CN108288033B (en) | A kind of safety cap detection method based on random fern fusion multiple features | |
| CN105844295B (en) | A kind of video smoke sophisticated category method based on color model and motion feature | |
| CN107909081B (en) | Method for quickly acquiring and quickly calibrating image data set in deep learning | |
| CN103605953B (en) | Vehicle interest target detection method based on sliding window search | |
| CN109145708B (en) | Pedestrian flow statistical method based on RGB and D information fusion | |
| CN108197587A (en) | A kind of method that multi-modal recognition of face is carried out by face depth prediction | |
| CN107507172A (en) | Merge the extra high voltage line insulator chain deep learning recognition methods of infrared visible ray | |
| CN104978567B (en) | Vehicle checking method based on scene classification | |
| CN107145845A (en) | The pedestrian detection method merged based on deep learning and multi-characteristic points | |
| CN102214309B (en) | Special human body recognition method based on head and shoulder model | |
| CN105046206B (en) | Based on the pedestrian detection method and device for moving prior information in video | |
| CN106127799B (en) | A kind of visual attention detection method for 3 D video | |
| CN106886216A (en) | Robot automatic tracking method and system based on RGBD Face datections | |
| CN110264493A (en) | A kind of multiple target object tracking method and device under motion state | |
| CN109045676B (en) | Chinese chess recognition learning algorithm and robot intelligent system and method based on algorithm | |
| CN106339657B (en) | Crop straw burning monitoring method based on monitor video, device | |
| JP2001195591A (en) | Method for detecting void in image | |
| JP2001202525A (en) | Method for deciding direction of image including blue sky |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170510 |
|
| RJ01 | Rejection of invention patent application after publication |