CN101520850B - Construction method of object detection classifier, object detection method and corresponding system - Google Patents
Construction method of object detection classifier, object detection method and corresponding system Download PDFInfo
- Publication number
- CN101520850B CN101520850B CN2009100820387A CN200910082038A CN101520850B CN 101520850 B CN101520850 B CN 101520850B CN 2009100820387 A CN2009100820387 A CN 2009100820387A CN 200910082038 A CN200910082038 A CN 200910082038A CN 101520850 B CN101520850 B CN 101520850B
- Authority
- CN
- China
- Prior art keywords
- characteristic
- band characteristic
- band
- image
- sorter
- 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.)
- Active
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a construction method of an object detection classifier, an object detection method and a corresponding system. The object detection method comprises the steps of constructing band characteristic which is formed by a plurality of alternate sub-bands, calculating the responding value of the band characteristic in a sample image, constructing a classifier according to the responding value, detecting the image by utilizing the classifier and obtaining an object detection result. The invention has the advantages that the band characteristic has stronger description ability for complex object shapes and has stronger robustness for translation and measure error to a certain extent.
Description
Technical field
The present invention relates to computer vision and area of pattern recognition, further relate to a kind of method and system that in image or video, carries out object detection.
Background technology
From image or video, detect other object of specified class and be one meaningful and be imbued with challenging work.In the prior art; Common way is that a kind of descriptive power of design is strong and calculate simple characteristic set; Utilize existing feature selection approach from the characteristic set that comprises big measure feature, to select the best character subset of discriminating power then such as adaptive boosting algorithm (AdaBoost); And with the latent structure sorter in the character subset of selecting, application scanning window scanning strategy and post-processing algorithm are carried out object detection.The common way of object detection is to describe the position and the scale size of object with a rectangle, and this rectangle makes object be positioned at its center, and the yardstick of this rectangle is identical or similar with the boundary rectangle of object size, claims that this rectangle is a scanning window.
Particularly, object detecting method comprises the following steps.At first, characteristic complete or collected works that include big measure feature of exhaustive generation in a scanning window.Then, calculate the response of the characteristic of sample in positive routine sample set and the counter-example sample set, adopt machine learning algorithm from characteristic set, to select " object " and " non-object " characteristic that discriminating power is the strongest according to above-mentioned response, and the structural classification device.Boosting algorithm (boosting) commonly used in feature selecting is comprising adaptive boosting algorithm (AdaBoost) and continuous boosting algorithm (RealBoost).The basic process of boosting algorithm work is to iterate continuously, and each iteration is selected the characteristic of an optimum and constructed a corresponding Weak Classifier.After iteration n time, obtain n characteristic and Weak Classifier respectively, these Weak Classifiers are further formed a cascade sort device, also are strong classifier.Fig. 1 representes the principle of work of cascade sort device.The image block of each input is differentiated by each Weak Classifier successively, is input as " non-object " if one of them Weak Classifier is judged, then this image block will directly be judged as " non-object "; Have only when n Weak Classifier all is judged to be " object " with this image block, this image block just finally is considered to " object ".
At last, detect most of scanning window method for scanning that adopts based on above-mentioned cascade sort device.Fig. 2 has explained the flow process of scanning window scanning.At first image is carried out scale; Utilize the scanning window (for example pixel is 64 * 32 scanning window) of unified size to scan to all locus in the image of each yardstick of obtaining after the conversion then; Calculate the response of each characteristic of the image block in each scanning window; And differentiate according to the response of this characteristic, obtain candidate window.Carry out aftertreatment then; Comprising utilizing the noise remove algorithm to remove obvious noise; And then utilize overlapping scan window merge algorithm (for example mean filter algorithm (MeanShift)) a plurality of scanning windows that same object is corresponding to merge, thus obtain final detection result---the object window.
In above-mentioned technology path, design fast and effectively that characteristic is a most important component, the characteristic that the structure description ability is strong and computation complexity is low is to realize a key factor of fast detecting.In the prior art, have two kinds of better characteristics: Lis Hartel is sought peace based on the characteristic of profile.Paul Viola and Michael Jones will describe gray scale and difference in the adjacent rectangle zone in the Rapid of the CVPR of calendar year 2001 Object Detection Using aBoosted Cascade of Simple Features one literary composition Lis Hartel is levied (haar-like feature) and is incorporated in the object detection.Lis Hartel is levied as shown in Figure 3, and Lis Hartel is levied it and is made up of several regular rectangles, the response of characteristic be shadow region interior pixel brightness and with the brightness of non-hatched area interior pixel and poor.During levying, Lis Hartel comprises the edge feature that reflects certain width---the limit characteristic, shown in the last figure of Fig. 3; Characteristic---the ridge characteristic of the crestal line of reflection certain width is shown in the following left figure of Fig. 3; And reflect certain specific light and shade zone characteristic of pattern alternately, shown in the following right figure of Fig. 3.In the rectangle of each above-mentioned rule all pixel intensity with can calculate through the integrogram fast algorithm.People's face that Lis Hartel is levied at fast robust obtains immense success in detecting and using.Bo Wu and Ram Nevatia have proposed a kind of little limit characteristic (edgelet) of describing simple profile information in the image at the Detection of ICCV in 2005 of Multiple in Partially Occluded Humans in a Single Image byBayesian Combination of Edgelet Part Detectors one literary composition.Little limit characteristic is as shown in Figure 4, and it is the segment of curve of some ad-hoc locations in the scanning window, comprises straight-line segment and 1/2,1/4,1/8 circular arc line segment (all not illustrating in the drawings) at any angle.When extracting the little limit of image characteristic, at first the gradient image of computed image obtains final response through the compute gradient image gradient similarity degree corresponding with little limit characteristic then.Little limit characteristic has obtained effect preferably in human detection and vehicle detection.J.Shotton; Also proposed a kind of in A.Blake and R.Cipolla Contour-BasedLearning for Object Detection one literary composition in ICCV in 2005 with the characteristic of outline line fragment as object detection; And carry out similarity with outline algorithm (chamfer) mode and calculate, in the experiment of vehicle and motorcycle, obtained good effect.Yet above-mentioned two kinds of characteristics all have deficiency separately; The pattern that Lis Hartel is levied is very simple; As can be seen from Figure 3; The pattern of its reflection only comprises level and vertical straight-line segment, thereby can obtain effect preferably for detecting the fairly simple object of people's face isotype, but for shape more complex objects then performance is not ideal enough; Though the characteristic based on the outline line design can reflect the shape facility that object is comparatively complicated; But it is a kind of characteristic of single pixel wide; If the position of edge in the image and characteristic have slightly the dislocation and yardstick inconsistent; The response of final characteristic just possibly will alter a great deal, and that is to say that the characteristic of outline line is too responsive to translation and change of scale.
Summary of the invention
The technical matters that the present invention will solve is the robust of complex object in image or the video, accurately and apace detects, thereby has proposed a kind of object detecting method and system.
According to an aspect of the present invention, a kind of building method that is used for the sorter of object detection is provided, has comprised the following steps:
110) structure band characteristic, wherein said band characteristic is made up of a plurality of alternate sub-bands;
120) calculate the response of said band characteristic in sample image;
130) according to said response structural classification device.
In the building method of this sorter, said step 110) further comprise:
111) parameter of the size of input scan window and said band characteristic;
112) confirm the outline line of said band characteristic according to the parameter of the size of said scanning window and said band characteristic;
113) expand a plurality of alternate sub-bands by said outline line, form said band characteristic by said sub-band.
In the building method of this sorter, said step 120) be to calculate the response of said band characteristic in sample image through mean luminance differences or the weighted mean luminance difference calculated between the said alternate sub-band.
In the building method of this sorter, said step 120) further comprise:
121) integrogram of the said sample image of calculating;
122) utilize the integrogram of said sample image to calculate the response of said band characteristic in sample image.
In the building method of this sorter, said step 122) for the situation that said band characteristic is not Ha Er, utilize said sub-band in connect positive rectangle and be similar to said sub-band.
In the building method of this sorter, said step 130) further comprise:
Calculate the discriminating power of said band characteristic and the false drop rate of current sorter according to said response, from said band characteristic, select preferred band characteristic according to the calculation cost of said discriminating power, said false drop rate and said band characteristic.
According to a further aspect in the invention, a kind of system that is configured to the sorter of object detection is provided also, has comprised characteristic generating apparatus, feature deriving means and sorter constructing apparatus, wherein:
Said characteristic generating apparatus is used to construct the band characteristic, and wherein said band characteristic is made up of a plurality of alternate sub-bands;
Said feature deriving means is used for receiving sample image and said band characteristic from said sorter constructing apparatus, and calculates the response of said band characteristic in said sample image;
The sorter constructing apparatus is used to receive sample image and receives said band characteristic from said characteristic generating apparatus, according to said response structural classification device.
According to another aspect of the invention, a kind of object detecting method is provided also, has comprised the following steps:
210) according to above-mentioned building method structural classification device;
220) use said sorter image is detected, obtain object detection result.
In this object detecting method, said step 220) further comprise:
221) treat detected image and carry out scale, and said scaled images is carried out window scanning;
222) response of the said band characteristic in the said window of said classifier calculated, and according to said response acquisition candidate window;
223) to said candidate window aftertreatment, obtain said object detection result.
According to another aspect of the invention, a kind of object detecting system is provided also, has comprised:
The window scanister is used to receive image to be detected, treats detected image and carries out scale and window scanning;
Feature deriving means is used to calculate the used response of band characteristic in said window of structural classification device;
Sorter according to above-mentioned building method structure is used for obtaining candidate window according to said response;
After-treatment device is used for said candidate window aftertreatment, obtains object detection result.
Effect of the present invention is that characteristic has stronger descriptive power to complex objects shape comparatively, for to a certain degree translation, scale error stronger robustness is arranged simultaneously.
Description of drawings
Do further detailed explanation below in conjunction with the accompanying drawing specific embodiments of the invention, wherein:
Fig. 1 is the waterfall type detection algorithm synoptic diagram of prior art;
Fig. 2 is the multiple dimensioned window scanning process process flow diagram of prior art;
Fig. 3 is that the Lis Hartel of prior art is levied synoptic diagram;
Fig. 4 is the little limit characteristic synoptic diagram of prior art;
Fig. 5 is the band characteristic synoptic diagram in the scanning window according to an embodiment of the invention;
Fig. 6 is the synoptic diagram of four kinds of typical band characteristics according to an embodiment of the invention;
Fig. 7 is that the band characteristic generates block diagram according to an embodiment of the invention;
Fig. 8 is a feature extraction block diagram according to a preferred embodiment of the invention;
Fig. 9 is the synoptic diagram that utilizes integrogram calculating band characteristic according to a preferred embodiment of the invention;
Figure 10 is that constructed object detects the system of sorter and the overall framework figure of object detecting system according to an embodiment of the invention;
Figure 11 is the part vehicle sample example schematic diagram of collecting according to an embodiment of the invention;
Figure 12 is the test result synoptic diagram on various visual angles vehicle testing data set according to an embodiment of the invention.
Embodiment
In order to make the object of the invention, technical scheme and advantage clearer, below in conjunction with accompanying drawing, to further explain of the present invention.Should be appreciated that specific embodiment described herein only in order to explanation the present invention, and be not used in qualification the present invention.
According to a specific embodiment of the present invention, provide a kind of and be used for carrying out the method for the sorter of object detection at image to be detected according to the band latent structure, wherein this band characteristic is made up of a plurality of adjacent sub-bands.
Structural map is as the band characteristic, and this band characteristic includes but not limited to linear pattern band at any angle, and 1/4,1/2,1/8 circular arc type band, also is that it can be the band of random geometry.Each band is made up of a plurality of adjacent sub-bands.
Image band Q-character is in scanning window, and Fig. 5 provides an image band characteristic that is arranged in scanning window, and wherein the bold dots line is its outline line.The shape S of this band feature contour line is a straight-line segment; Outline line length is L, and the starting point of outline line is positioned at the p point of scanning window, and the angle of outline line or its starting point place tangent line and horizontal direction is θ; The sub-band number t of this band characteristic is 3, and the width of these three brief note bands is respectively w
1, w
2, w
3, the band characteristic can parameter turn to thus
Can be through control S, L, p, θ, t, { w
i}
I=1 tIsoparametric span and step-length, the band characteristic set of generation different scales.Listed four kinds of typical band characteristics according to an embodiment of the invention among Fig. 6, linear pattern limit characteristic, linear pattern ridge characteristic, arc line type limit characteristic and arc line type ridge characteristic, the sub-band that the band characteristic shown in the figure is equated by width is formed alternately.One of ordinary skill in the art will appreciate that each sub-strip width w
RAlso can get different widths.
The band characteristic that Fig. 7 shows according to the specific embodiment of the invention generates block diagram; Construct above-mentioned image strip band characteristic and comprise the following steps: the at first size of input scan window and the parameter of band characteristic; Wherein not too much for the number of features among the controlling features complete or collected works, usually these parameters are limited, concrete; For example the scanning window restriction L to 64 * 32 pixel sizes is 4 to 12 pixels, w
RBe 2 to 6 pixels.Then, confirm the outline line of each band characteristic, outline line can be the curve of arbitrary shape, and straight-line segment that adopts in the embodiments of the invention and segment of curve are the example of band feature contour line.In scanning window, select the position p of band characteristic, the position p of band characteristic can represent that it can be rational arbitrarily position in the scanning window with the starting point of outline line.Angle through the band characteristic and shape are calculated another end points of band feature contour line, and to generate length according to the geometrical constraint interpolation of straight line or camber line be the outline line of L.At last, outline line is extended to each image band characteristic, it is several sub-band combination.Expand same width w along normal direction to both sides by outline line
R, generate the limit characteristic; Expand same width 1/2w along normal direction to both sides with the reason outline line
R, further outwards expand w respectively again
RWidth generates the ridge characteristic.Each point on the outline line after the record expansion and all parameter S of band characteristic, L, p, θ, t, { w
i}
I=1 t, promptly constructed the band characteristic.One of ordinary skill in the art will appreciate that, can adopt multiple computer graphics method to confirm outline line, equally also can adopt multiple graphics method to come to generate image band characteristic from outline line.
Calculate the response of each image band characteristic in sample image.The essence of band characteristic is to be used for weighing the contrast situation of the brightness of non-shaded bar region and shaded bar region among Fig. 6, so can be regarded as the response account form of image band characteristic so long as reflect the computing method of alternate sub-band brightness contrast situation.Specific embodiment according to the present invention, the absolute value of difference of mean flow rate that has adopted non-hatched area and shadow region is as the response of characteristic, and the mode of extraction can be expressed as with formula:
Wherein I (x, y) denotation coordination (x, the pixel value of y) locating, the sub-banded zone in R1, R2, the R3 presentation graphs 1, || || the number of a regional interior pixel is asked in expression.
More general, can give different weights to the brightness of each point in the band, concrete formula can be expressed as formula (3) and (4):
Symbol in formula (3), (4) is consistent with the symbol in formula (1), (2), and wherein (x is that (x y) locates corresponding weights to some y) to q.When weights all in formula (3), (4) all were 1, formula (3) (4) just deteriorated to formula (1), (2).
One of ordinary skill in the art will appreciate that; The feature extraction mode of above-mentioned formula (1) to (4) representative is a kind of possible implementation of image band characteristic; In addition the alternate manner that also has the sub-band light and shade contrast of a lot of reflections; For example the absolute value sign in the formula (1) to (4) is removed, also can be used as a kind of implementation of image band feature extraction.
The response of top computed image band characteristic is direct calculation mode, and the point that in image band characteristic area, comprises is a lot, and at this time calculated amount will be very big.According to a preferred embodiment of the invention, a kind of method of response of quick calculated characteristics is provided also, has adopted a plurality of rectangles to be similar to each sub-band, thereby can effectively utilize integrogram to calculate the response of band characteristic.
Block diagram as shown in Figure 8 for the image band characteristic of an appointment, at first judges whether it is that Lis Hartel is levied, and then adopts integrogram directly to calculate if Lis Hartel is levied, and levies employing approximation method calculated response value for non-Lis Hartel.Particularly, at first, calculate the integrogram of sample image, calculate the response of characteristic in the scanning window then.If the outline line of image band characteristic is level or vertical straight line, promptly this band characteristic is positive rectangle, and this type band characteristic is exactly that Lis Hartel is levied so, can directly adopt based on the quick Lis Hartel of integrogram and levy method for distilling.Levy for other non-Lis Hartel, adopt approximate mode shown in Figure 9 to extract characteristic.For the limit characteristic, shown in Fig. 9 (A) and Fig. 9 (C), calculate each the pixel P on the outline line
iThe intersection point A of outline of normal and band characteristic
iAnd B
i, respectively with line segment P
iA
iAnd P
iB
iConfirm the rectangle R1 of two rules for diagonal line
i, R2
iFor the every bit on the outline line, can determine up and down in two in two sub-bands and connect positive rectangle.Calculating through to every bit on the outline line can obtain two groups of positive rectangles, is respectively applied for approximate two sub-bands up and down.One of ordinary skill in the art will appreciate that the foregoing description has only provided the building method that connects positive rectangle in a kind of, can also adopt in the alternate manner structure and connect positive rectangle.Pixel in each rectangle with can utilize the integrogram algorithm to calculate fast, final feature deriving means for the computing formula of limit characteristic is:
Wherein g () expression utilize the integrogram fast algorithm calculate all pixels in the zone and, || || the same with the implication in formula (1), (2), represent the number of all pixels in the zone.
The calculating of ridge characteristic is similar with the calculating of above-mentioned limit characteristic, shown in Fig. 9 (B) and 9 (D), for each point on the outline line, still asks the intersection point of its normal and each sub-band outline, can confirm 4 some A like this
i, B
i, C
iAnd D
i, like this with A
iB
i, B
iC
i, C
iD
iFor diagonal line can confirm to connect in three positive rectangle R1
i, R2
i, R3
i, the computing formula of eigenwert is following:
Because the rectangle of above-mentioned approximate band only depends on the shape of image band characteristic, and irrelevant with input picture, therefore, preferably off-line ground obtains the parameter of each rectangle.Directly utilize integrogram to carry out feature extraction then, thereby improve execution speed according to these parameters.
From the characteristic complete or collected works that comprise big measure feature, select current positive routine sample set and the optimum character subset of counter-example sample set according to the response of the characteristic of being calculated; And utilize optimum character subset to construct Weak Classifier, further by Weak Classifier structure cascade sort device.One of ordinary skill in the art will appreciate that, can adopt any suitable learning algorithm to carry out feature selecting, like adaptive boosting algorithm and continuous boosting algorithm or the like, be its principle of work of example explanation with continuous boosting algorithm only here.At first; For the statistic histogram distribution respectively of the response of each characteristic on the sample set of positive routine sample set and counter-example; Ask the Bart of the response histogram distribution of this characteristic on all positive examples and counter-example sample to look into subbreed number (Bhattacharyya Coefficient) then, formulate is following:
W wherein
+ jRepresent this characteristic j component in the response histogram on positive routine sample set, W
- jRepresent this characteristic j component in the response histogram on the counter-example sample set, this characteristic of the more little expression of Z value of certain characteristic is strong more for positive example and counter-example discriminating power, promptly should preferentially select this characteristic.
In order to select image band characteristic better more effectively to classify; Preferably; The present invention is except the factor of the discriminating power of having considered the band characteristic; Also from optimizing feature selecting when the false drop rate of the sorter of front construction and/or calculation cost two aspects of band characteristic, the above-mentioned implementation of having weighed discriminating power, false drop rate and calculation cost is referred to as the characteristic balance.Particularly; At first obtain strong classifier that current study obtains false drop rate fp; Can also can obtain through estimating, for example directly from continuous boosting algorithm, obtain the estimated value of false drop rate at the current sorter of the set operation of counter-example sample image through multiple alternate manner.According to the discriminating power of characteristic, the calculation cost C of characteristic and the fp of the strong classifier that current study obtains, the band characteristic is weighed calculating according to formula (8):
Wherein α is a weighting coefficient, and according to one embodiment of present invention, α is positioned at (0,1) interval; W
+ j, W
- jIdentical with the physical significance of representative in the formula (7); Fp is the false drop rate of the strong classifier that obtains of current study; C is a calculation cost, and it is according to the computation complexity of each characteristic and concrete implementation method calculated off-line.Explain formula (8) intuitively, the false drop rate of system is bigger at the beginning, and system tends to the characteristic of selecting calculation cost little; And along with the false drop rate of the system of going deep into of training diminishes gradually, system is more prone to the characteristic of selecting discriminating power strong, thereby reaches adaptive adjusting.Table 1 has provided the one group of calculation cost that all kinds of band characteristics are corresponding that is embodied as example with CPU.Can estimate according to concrete system architecture (like the shared clock period of multiplication and addition etc.) for the realization of different platform, perhaps confirm calculation cost through experiment.
Table 1
Preferably, according to the characteristic balance result of formula (8), that characteristic of selecting the minimum Z ' value of correspondence is as current optimal characteristics.Should be appreciated that, top explanation just with continuous boosting algorithm as an instance, but the step of characteristic balance can be used with the multiple machine learning algorithm collocation that is used for feature selecting.
The scanning window that utilizes selected characteristic to be configured to differentiate input is the Weak Classifier of " object " or " non-object ".By the common structure of a plurality of Weak Classifier series connection strong classifier, cascade sort device as shown in Figure 3.At last, optimal feature subset, Weak Classifier and the reference record of cascade sort device of the cascade sort device that obtains are got off.
According to a specific embodiment of the present invention, the present invention also provides a kind of cascade sort device of using above-mentioned structure to carry out the method for the object detection in the image.Image to be detected can be the image of gray scale, colour, infrared or other type.Preferably, before feature extraction, be converted into luminance picture or other can reflect the image of object band characteristic.According to a preferred embodiment of the invention, at first treat detected image and carry out scale, and scaled images is carried out window scanning.Calculate the response of the band characteristic of cascade sort device, and carry out object detection according to this response at scanning window.This object detection result is carried out aftertreatment again; According to one embodiment of present invention; Aftertreatment comprises: at first utilize the noise remove method to remove the scanning window that obviously is not object, noise remove can pass through accomplished in many ways, and for example non-maximal value suppresses algorithm; And then utilize heavy frame merging method to be merged into final testing result to unnecessary heavy frame.Preferably, above-mentioned response according to the integrogram of scaled images or convergent-divergent after the integrogram of window of image calculate.
According to specific embodiment shown in figure 10, the present invention also provides a kind of system and a kind of object detecting system that is configured to the sorter of object detection.
Specific embodiment according to the present invention, this system that is configured to the sorter of object detection comprises: characteristic generating apparatus, feature deriving means and sorter constructing apparatus, wherein:
The characteristic generating apparatus is used to construct the band characteristic, and wherein said band characteristic is made up of a plurality of alternate sub-bands, and band characteristic complete or collected works are sent to the sorter constructing apparatus.
Feature deriving means is used for receiving band characteristic complete or collected works and sample image from the sorter constructing apparatus, and calculates the response of said band characteristic in said sample image.Wherein, said sample image comprises the image block of a large amount of positive examples, for example 3000~5000.Figure 11 shows the part vehicle sample example according to the used collection of training cascade sort device of a specific embodiment of the present invention.These samples all are according to the intercepting from the image that comprises vehicle of prior markup information, then all samples are normalized to 64 * 32 pixel sizes, and make vehicle all be positioned at the center of image block.
The sorter constructing apparatus; Receiving positive routine sample set and counter-example sample set merges from characteristic generating apparatus reception band characteristic complete or collected works; From the characteristic complete or collected works that comprise a large amount of band characteristics, select current positive routine sample set and the optimum character subset of counter-example sample set according to said characteristic response value; And utilize optimum character subset to construct Weak Classifier, further by Weak Classifier structure cascade sort device.According to the preferred embodiment of the present invention, at the process synthesis of feature selecting the factor of false drop rate of calculation cost and sorter of band characteristic.
In accordance with a preferred embodiment of the present invention, above-mentioned object detecting system comprises:
The window scanister is used to receive image to be detected, and said image to be detected is carried out scale, and said scaled images is carried out window scanning;
Feature deriving means is used for calculating the response of band characteristic at said window; Identical with the feature deriving means that sorter generates part at aspects such as principle, building methods, difference only is that the image band characteristic of its input is the optimal characteristics that chooses through the feature selecting device.
The sorter of being constructed as stated is used for obtaining candidate window according to said response;
After-treatment device is used for said candidate window is carried out aftertreatment, obtains object detection result.
The vehicle detection result of a specific embodiment according to the present invention to various visual angles has been shown among Figure 12, and transverse axis is represented accuracy rate, and the longitudinal axis is represented recall rate, and curve high-performance more is good more.Can see that the testing result of the present invention on the various visual angles automobile obviously is better than utilizes Lis Hartel to seek peace little limit characteristic to the testing result of various visual angles automobile, so the present invention has stronger robustness for to a certain degree translation, scale error.Experiment also shows, through adopt approximate treatment banded zone interior pixels with---adopt a plurality of positive rectangles be similar to each sub-band, and then the employing integrogram, carry out feature extraction fast, effectively improved computing velocity, and saved storage space; And adopted the object detecting system that the characteristic balance constructs can be than the detection system that does not adopt the characteristic balance to construct fast 2 to 3 times, and performance be constant basically.
It is to be platform with the PC that said system realizes, but the present invention can have multiple concrete platform, includes but not limited to PC, server, SoC, FPGA or the like.
Should be noted that and understand, under the situation that does not break away from the desired the spirit and scope of the present invention of accompanying Claim, can make various modifications and improvement the present invention of above-mentioned detailed description.Therefore, the scope of the technical scheme of requirement protection does not receive the restriction of given any specific exemplary teachings.
Claims (9)
1. a building method that is used for the sorter of object detection comprises the following steps:
110) parameter of the size of input scan window and band characteristic; Confirm the outline line of said band characteristic according to the parameter of the size of said scanning window and said band characteristic; Expand a plurality of alternate sub-bands by said outline line, form said band characteristic by said a plurality of alternate sub-bands;
120) through calculating mean luminance differences or the weighted mean luminance difference between the said alternate sub-band or using the integrogram of said sample image to calculate the response of said band characteristic in sample image;
130) according to said response structural classification device.
2. according to the building method of the said sorter of claim 1; It is characterized in that; When using the integrogram of said sample image to calculate the response of said band characteristic in sample image said step 120); For the situation that said band characteristic is not Ha Er, utilize said sub-band in connect positive rectangle and be similar to said sub-band.
3. the building method of sorter according to claim 1 is characterized in that, said step 130) further comprise:
Calculate the discriminating power of said band characteristic and the false drop rate of current sorter according to said response, from said band characteristic, select preferred band characteristic according to the calculation cost of said discriminating power, said false drop rate and this band characteristic.
4. according to the building method of the said sorter of claim 1, it is characterized in that the width of said sub-band equates.
5. according to the building method of the said sorter of claim 1, it is characterized in that said outline line is straight-line segment or 1/2,1/4,1/8 circular arc line segment at any angle.
6. a system that is configured to the sorter of object detection comprises characteristic generating apparatus, feature deriving means and sorter constructing apparatus, wherein:
Said characteristic generating apparatus; Be used for the size of input scan window and the parameter of band characteristic; Confirm the outline line of said band characteristic according to the parameter of the size of said scanning window and said band characteristic; Expand a plurality of alternate sub-bands by said outline line, form said band characteristic by said a plurality of alternate sub-bands;
Said feature deriving means; Be used for receiving sample image and said band characteristic, and calculate the response of said band characteristic in said sample image through mean luminance differences or the weighted mean luminance difference calculated between the said alternate sub-band from said sorter constructing apparatus;
The sorter constructing apparatus is used to receive sample image and receives said band characteristic from said characteristic generating apparatus, according to said response structural classification device.
7. an object detecting method comprises the following steps:
210) building method structural classification device according to claim 1;
220) use said sorter image is detected, obtain object detection result.
8. according to the said object detecting method of claim 7, it is characterized in that said step 220) further comprise:
221) treat detected image and carry out scale, and the image after the said scale is carried out window scanning;
222) response of the said band characteristic in the said window of said classifier calculated, and according to the said response acquisition candidate window of band characteristic in image to be detected;
223) to said candidate window aftertreatment, obtain said object detection result.
9. object detecting system comprises:
The window scanister is used to receive image to be detected, treats detected image and carries out scale and window scanning;
Feature deriving means is used to calculate the used response of band characteristic in said window of structural classification device;
The sorter of building method structure according to claim 1 is used for obtaining candidate window according to said band characteristic in response described in the image to be detected;
After-treatment device is used for said candidate window aftertreatment, obtains object detection result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN2009100820387A CN101520850B (en) | 2009-04-17 | 2009-04-17 | Construction method of object detection classifier, object detection method and corresponding system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN2009100820387A CN101520850B (en) | 2009-04-17 | 2009-04-17 | Construction method of object detection classifier, object detection method and corresponding system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN101520850A CN101520850A (en) | 2009-09-02 |
CN101520850B true CN101520850B (en) | 2012-02-15 |
Family
ID=41081431
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN2009100820387A Active CN101520850B (en) | 2009-04-17 | 2009-04-17 | Construction method of object detection classifier, object detection method and corresponding system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN101520850B (en) |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103793722A (en) * | 2014-03-07 | 2014-05-14 | 电子科技大学 | Rapid vehicle detection method and device under low resolution |
CN106575364A (en) * | 2014-04-11 | 2017-04-19 | 英特尔公司 | Object detection using directional filtering |
US9697443B2 (en) | 2014-12-11 | 2017-07-04 | Intel Corporation | Model compression in binary coded image based object detection |
CN106022235B (en) * | 2016-05-13 | 2021-05-28 | 中国人民解放军国防科学技术大学 | Missing child detection method based on human body detection |
JP6635074B2 (en) * | 2017-03-02 | 2020-01-22 | オムロン株式会社 | Watching support system and control method thereof |
CN113192358A (en) * | 2021-04-26 | 2021-07-30 | 贵州车秘科技有限公司 | Parking management system based on thermal imaging technology in intelligent parking field and use method thereof |
-
2009
- 2009-04-17 CN CN2009100820387A patent/CN101520850B/en active Active
Non-Patent Citations (3)
Title |
---|
陈勇飞等.基于肤色和类Harr特征的人脸图像的人眼检测.《计算机工程与应用》.2008,(第33期),174-176,180. * |
黄雄栋.人脸识别***中的特征检测算法研究.《中国优秀硕士学位论文全文数据库》.2009,(第3期),10-20. * |
黄雄栋.人脸识别***中的特征检测算法研究.《中国优秀硕士学位论文全文数据库》.2009,(第3期),第10-20页. * |
Also Published As
Publication number | Publication date |
---|---|
CN101520850A (en) | 2009-09-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106874894B (en) | Human body target detection method based on regional full convolution neural network | |
JP6547069B2 (en) | Convolutional Neural Network with Subcategory Recognition Function for Object Detection | |
CN102722712B (en) | Multiple-scale high-resolution image object detection method based on continuity | |
CN101350063B (en) | Method and apparatus for locating human face characteristic point | |
EP2434431A1 (en) | Method and device for classifying image | |
CN101520850B (en) | Construction method of object detection classifier, object detection method and corresponding system | |
CN108681693B (en) | License plate recognition method based on trusted area | |
CN102609716B (en) | Pedestrian detecting method based on improved HOG feature and PCA (Principal Component Analysis) | |
CN111640089B (en) | Defect detection method and device based on feature map center point | |
CN101561867B (en) | Human body detection method based on Gauss shape feature | |
CN101178770B (en) | Image detection method and apparatus | |
CN101655914B (en) | Training device, training method and detection method | |
US8363983B2 (en) | Real-time face detection apparatus | |
CN104766046A (en) | Detection and recognition algorithm conducted by means of traffic sign color and shape features | |
CN104809452A (en) | Fingerprint identification method | |
CN100561501C (en) | A kind of image detecting method and device | |
CN106056751A (en) | Prefix number identification method and system | |
CN101930549B (en) | Second generation curvelet transform-based static human detection method | |
CN109376740A (en) | A kind of water gauge reading detection method based on video | |
CN100561505C (en) | A kind of image detecting method and device | |
EP2648159A1 (en) | Object detecting method and object detecting device using same | |
CN114998852A (en) | Intelligent detection method for road pavement diseases based on deep learning | |
CN101975940A (en) | Segmentation combination-based adaptive constant false alarm rate target detection method for SAR image | |
CN102867195B (en) | Method for detecting and identifying a plurality of types of objects in remote sensing image | |
CN104036284A (en) | Adaboost algorithm based multi-scale pedestrian detection method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant |