CN101719276B - The method and apparatus of object in a kind of detected image - Google Patents
The method and apparatus of object in a kind of detected image Download PDFInfo
- Publication number
- CN101719276B CN101719276B CN200910238688.6A CN200910238688A CN101719276B CN 101719276 B CN101719276 B CN 101719276B CN 200910238688 A CN200910238688 A CN 200910238688A CN 101719276 B CN101719276 B CN 101719276B
- Authority
- CN
- China
- Prior art keywords
- image
- search
- scaling
- sorter model
- hunting zone
- 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
- 238000000034 method Methods 0.000 title claims abstract description 60
- 238000012545 processing Methods 0.000 claims abstract description 14
- 238000001514 detection method Methods 0.000 claims description 18
- 238000012360 testing method Methods 0.000 claims description 16
- 230000009467 reduction Effects 0.000 claims description 4
- 230000008569 process Effects 0.000 description 24
- 238000010586 diagram Methods 0.000 description 7
- 238000005516 engineering process Methods 0.000 description 7
- 238000012549 training Methods 0.000 description 5
- 230000008901 benefit Effects 0.000 description 3
- 238000000605 extraction Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Abstract
The invention provides the method and apparatus of object in a kind of detected image, method comprises: step one, calibrating camera, obtains the corresponding relation between the shooting pixels tall of object target in the shooting image of described video camera and the true altitude of object target; Step 2, arranges mask images, arranges the mask pixels height of object target in mask images according to described true altitude; Step 3, obtains the sorter model of described object target according to mask images; Step 4, the mode of scaling is adopted to obtain the scaling image of the different scale of described shooting image, according to described mask pixels height, described true altitude and described corresponding relation, determine the region of search of described sorter model in described scaling image, and utilize described sorter model to search in described region of search.The present invention reduces by limiting target search yardstick the redundancy window that will search for, and improves search speed, solves the technical matters that prior art processing speed is very slow.
Description
Technical field
The present invention relates to image processing techniques, particularly relate to the method and apparatus of object in the detected image based on camera calibration technology.
Background technology
Object detection is the basis of video analysis, has important using value.Counterbody is followed the tracks of and foreground extraction matching technique, and object detection technology has more outstanding robustness (robustness, robustness) for illumination and noise.But existing object detection technology, often in order to the target of various yardstick be detected, needs to carry out traversal search to the image of multiple yardstick, so processing speed is very slow, cannot adopt in many applications.
Summary of the invention
The object of this invention is to provide the method and apparatus of object in a kind of detected image, reducing by limiting target search yardstick the redundancy window that will search for, improving search speed, solving the technical matters that prior art processing speed is very slow.
To achieve these goals, on the one hand, provide the method for object in a kind of detected image, comprise the steps:
Step one, calibrating camera, obtains the corresponding relation between the shooting pixels tall of object target in the shooting image of described video camera and the true altitude of described object target;
Step 2, arranges mask images, arranges the mask pixels height of described object target in described mask images according to described true altitude;
Step 3, obtains the sorter model of described object target according to described mask images;
Step 4, the mode of scaling is adopted to obtain the scaling image of the different scale of described shooting image, according to described mask pixels height, described true altitude and described corresponding relation, determine the region of search of described sorter model in described scaling image, and utilize described sorter model to search in described region of search.
Preferably, in above-mentioned method, in described step 3, described sorter model is the yardstick by obtaining based on the feature extracting method of window is the image of M × N, and the yardstick of described object target in described sorter model is M ' × N '.
Preferably, in above-mentioned method, also comprise, setting M=M ', N=N '.
Preferably, in above-mentioned method, in described step 4, the scaling number of plies scope of carrying out described scaling is: (
), wherein,
For the minimum value of the number of plies,
For the maximal value of the number of plies.Function f loor (f) is less than the maximum integer of floating number f for getting, ceiling (f) is greater than the smallest positive integral of floating number f for getting, and scale is scaling factor and constant for being greater than 1, h
minfor the minimum value of described mask pixels height, h
maxfor the maximal value of described mask pixels height.
Preferably, in above-mentioned method, in described step 4, s=s is respectively to the number of plies
min, s
min+ 1 ... s
maxthe described scaling image of different scale be handled as follows respectively:
To the described scaling image of current layer number, the height hunting zone setting the pixels tall of described object target is: (N
s min, N
s max), wherein,
According to described region of search and described height hunting zone, determine effective hunting zone R that current layer number is corresponding
s;
By described hunting zone R
sinterior image zooming is original scale
comparison chart picture doubly, utilizes described sorter model to carry out search comparison in described comparison chart picture, determines that the image-region matched with described sorter model is as testing result;
Described testing result scaling is returned original scale, carries out skew reduction, and merge adjacent testing result.
Preferably, in above-mentioned method, determine effective hunting zone R that current layer number is corresponding
sprocess specific as follows:
In described region of search, search meets
Position and record the minimum value l of the horizontal ordinate of correspondence position
s, horizontal ordinate maximal value r
s, ordinate minimum value be t
s, ordinate maximal value b
s;
Described effective hunting zone is:
Wherein,
For the left side horizontal ordinate of described effective hunting zone,
For the top ordinate of described effective hunting zone,
For the right horizontal ordinate of described effective hunting zone,
For the following ordinate of described effective hunting zone.
Preferably, in above-mentioned method, also comprise, Gauss model is set according to current layer number, utilize described Gauss model to obtain and detect degree of confidence, get rid of false-alarm according to the pixels tall of described testing result and described detection degree of confidence.
The present invention also provides the device of object in a kind of detected image, comprising:
Demarcating module, for calibrating camera, obtains the corresponding relation between the shooting pixels tall of object target in the shooting image of described video camera and the true altitude of described object target;
Mask images arranges module, for: mask images is set, the mask pixels height of described object target in described mask images is set according to described true altitude;
Sorter model module, for: the sorter model obtaining described object target according to described mask images;
Search module, for: adopt the mode of scaling to obtain the scaling image of the different scale of described shooting image, according to described mask pixels height, described true altitude and described corresponding relation, determine the region of search of described sorter model in described scaling image, and utilize described sorter model to search in described region of search.
Preferably, in above-mentioned device, also comprise:
Merge module, for: testing result scaling is returned original scale, carries out skew reduction, and merge adjacent testing result.
Preferably, in above-mentioned device, also comprise:
False-alarm processing module, for: Gauss model is set according to current layer number, utilizes described Gauss model to obtain and detect degree of confidence, get rid of false-alarm according to the pixels tall of described testing result and described detection degree of confidence.
At least there is following technique effect in the embodiment of the present invention:
1) because the embodiment of the present invention have employed the demarcation information of video camera, the pixel size of target on image can be obtained, thus determine the search yardstick of target accordingly, make to need redundancy window to be processed to greatly reduce, not only decrease operand, improve the speed of object detection, decrease interference simultaneously.
2) utilize Gauss model to carry out false-alarm process, decrease false alarm.
3) the present invention is that object detection technology provides possibility in the application of embedded system and SOC hardware system.
Accompanying drawing explanation
The flow chart of steps of the method that Fig. 1 provides for the embodiment of the present invention;
The structural drawing of the device that Fig. 2 provides for the embodiment of the present invention.
Embodiment
For making the object of the embodiment of the present invention, technical scheme and advantage clearly, below in conjunction with accompanying drawing, specific embodiment is described in detail.
The ripe object detecting method traveled through based on pyramid image zooming and window.The step of method is described below:
First, process image zooming is obtained the image of a series of different scale, the fixing multiple of yardstick difference between adjacent layer image, and rear layer is obtained through image zooming by front layer.This multiple is called scaling factor; Such as, ground floor is set as and processes image size and the identical image of pixel intensity, second layer image be ground floor image through the yardstick that scaling obtains be ground floor image scale image doubly, scale can get 0.8, third layer image be second layer image through the yardstick that scaling obtains be second layer image scale image doubly, the like.
Then, for every tomographic image, adopt the sorter model of fixed size to judge each position in this tomographic image, see that whether it is by model, if passed through, then record position and size;
Finally, merging the frame by detecting of adjacent position, obtaining final object space and size.
The defect of said method is, for each position in the image of different levels, all need to adopt sorter model to judge, operand is very large, and because positional is too many, easily produces more interference, meanwhile, make sorter model more complicated.
The present invention adopts camera calibration technology, obtains the dimensional information of each position target in image, is limited searching image level and the scope of target, thus make in object detection process by this dimensional information, needs redundancy window to be processed to greatly reduce.While raising object detection speed, also reduce object interference.The inventive method can be applied with existing most of object detection combine with technique, has good actual application value.
The flow chart of steps of the method that Fig. 1 provides for the embodiment of the present invention, as shown in the figure, in detected image, the method for object comprises:
The first step 101, calibrating camera, obtains the corresponding relation between the shooting pixels tall of object target in the shooting image of described video camera and the true altitude of described object target;
Second step 102, arranges mask images, arranges the mask pixels height of described object target in described mask images according to described true altitude; (be specially: determine inspected object, and according to the true altitude of object and the relation between shooting pixels tall and the true altitude of described object, determine the pixels tall image of each position object in image; )
3rd step 103, obtains the sorter model of described object target according to described mask images; (be specially: the object classification device model of a training pre-dimensioning; )
4th step 104, the mode of scaling is adopted to obtain the scaling image of the different scale of described shooting image, according to described mask pixels height, described true altitude and described corresponding relation, determine the region of search of described sorter model in described scaling image, and utilize described sorter model to search in described region of search.(be specifically as follows: according to described pixels tall image, described sorter model, determine pyramid image parameter, and obtain the hunting zone of each layer of pyramid, and in described hunting zone, adopt object classification device model to obtain candidate target region.)
Below each step is further described:
The first step, calibrating camera, obtains the pixels tall of target object in image and the transformation relation of object true altitude; Concrete grammar, can adopt the same object of demarcation at least three place known altitude, and try to achieve the line equation that goes out and obtain, camera calibration is designated as prior art, and detail please refer to pertinent literature;
Second step, target setting object type also obtains the true altitude of target setting object, initialization one and the pixels tall image processing the sizes such as image, be set to the pixels tall of the target object of bottom, and be recorded in described pixels tall image corresponding position with everybody in computing image;
Certainly, also can adopt the pixel wide of scaling method determination intended target at image diverse location place, the correspondence image obtained is object pixel width images, and for subsequent treatment.
In addition, the pixels tall of the object that each point is top in mask can also be calculated as mask images value; Embodiment illustrates to record the pixels tall that each point is top below.
Further, an above-mentioned pixels tall can be calculated for each zonule, thus reduce EMS memory occupation amount, improve acquisition speed, be about to process image and be divided into a lot of zonule, each zonule comprises multiple pixel, for each zonule calculates the pixels tall value of a target object; The benefit done like this is, reduce the number of the pixels tall value of record, but precision is less better.
3rd step, the object classification device model of a training pre-dimensioning;
Assuming that object classification device model width is W
o, be highly H
othen extract subject image region and scaling is that pre-dimensioning is as positive sample, extract the image-region that do not comprise object and scaling be pre-dimensioning as anti-sample, extract the feature of positive sample and anti-sample, and adopt sorter training pattern to train the sorter model obtaining object.The extracting mode of feature has a lot, for human body, the better method of vehicle is for adopting gradient orientation histogram feature (HOG, histogram of Oriented Gradient), the mode of sorter training pattern also has a lot, can adopt preferably based on support vector machine (SVM) or self-adaptive enhancement algorithm (AdaBoost).
4th step, according to described pixels tall image, determines the hunting zone of each layer of pyramid image, and adopts predtermined category device model to judge each position;
Due to the object detection framework based on window, in order to detect that difference varies in size the object of position, need to adopt image zooming mode to obtain pyramid picture structure, and adopt the object classification device model of the pre-dimensioning trained to carry out traversal processing to the at all levels of pyramid diagram picture.If all carry out traversal search to every one deck of pyramid diagram picture, then need position to be processed a lot, operand is very large.
For pyramid diagram picture, assuming that altogether containing the image of N number of yardstick, be respectively the 0th, 1 ..N-1 layer, assuming that the 0th layer of scaling ratio relatively processing image is S
0, the scaling ratio of (n+1)th layer of relative n-th layer and scaling factor are scale, then n-th layer processes the scaling ratio of image is relatively S
n=S
0* scale
n, n=0,1..N-1.Assuming that the width of process image and highly be W and H, then n-th layer image width be highly W
nand H
n, can W be got
n=Round (S
n* W) and H
n=Round (S
n* H), wherein Round () is the computing that rounds up.The each position of the sorter model of pre-dimensioning to n-th layer is adopted to judge, the width of the detection target obtained in process image and the 1/S being highly sorter model width and height
n, assuming that object classification device model width is W
o, be highly H
o, then the target that obtains width and be highly W in process image is detected
o/ S
n, H
o/ S
n.
In the present invention, the above-mentioned sorter model of employing is called object detection to the process that image-region judges.
Visible, n-th layer can only detect process image in pixel wide and highly be W
o/ S
n, H
o/ S
ntarget.But the pixel wide of the target of the predefined type that diverse location place exists and highly different and different with position in n-th layer image.If all adopt predetermined Scale Model to judge to all positions, then can there is the redundant operation that the position of the target that can not there is described pixel size is also processed.
A kind of more excellent mode is:
4.1 obtain the parameter building pyramid diagram picture;
In order to pyramid image can be obtained from process image, need to know that the 0th layer, pyramid processes the scaling ratio of image relatively, and scaling factor, and determine the number of plies of pyramid diagram picture.Prior art can be adopted to determine.Preferably mode determines that the 0th layer, described pyramid processes the scaling yardstick of image relatively according to the maximum target height in described pixels tall image and/or width.And determine pyramidal total number of plies according to the scaling factor between adjacent layer further.
Illustrate for the pixels tall of target below.Assuming that in object pixel height image, the pixels tall of object is minimum is h
min, be h to the maximum
max, assuming that the scaling factor of image is scale, the ratio namely between adjacent image yardstick to be scale, scale be less than 1 constant, preferably scale can get 0.8.Then obtaining the 0th layer of scaling yardstick relatively processing image is
total number of plies of pyramid diagram picture is
because the number of plies is integer, consider and all yardsticks can be detected, can get total number of plies is
function Ceiling (f) is greater than the smallest positive integral of floating number f for getting.Further, the 0th layer of scaling yardstick relatively processing image can be limited and be not less than 1.Certainly, also can according to the minimum and maximum pixels tall of target of user's setting or the described parameter of width and scaling factor determination pyramid diagram picture.4.2 couples of n=0,1...N-1 are handled as follows respectively:
4.2.1 effective hunting zone that current layer number is corresponding is determined;
For current layer number, fixed size sorter is adopted to judge each position, the target width obtained and be highly W
o/ S
n, H
o/ S
n.Then in order to reduce operand, for current layer, search package is only needed to contain the position of current layer correspondence search yardstick, and for homologue volumetric pixel height not current layer scope position, then can not search for, like this, while minimizing operand, reduce flase drop.The embodiment defining efficient search scope can be as follows:
Setting search scope is a rectangle R
n(l
n, t
n, r
n, b
n), be set as that rectangle is that conveniently program processes.Assuming that for the position (x, y) in pixels tall image, the pixels tall of the target object of its correspondence is h (x, y).Set the altitude range of search target corresponding to current layer as (H
n min, H
n max), in process image (0,0, W, H) scope, search meets
Position, and the minimum value recording correspondence position horizontal ordinate is l
n, the maximal value r of horizontal ordinate
n, the minimum value of ordinate is t
n, the maximal value b of ordinate
n.(the H of each layer
n min, H
n max) establishing method be set each (H
n min, H
n max) combined covering (h
min, h
max) scope, and comprise H
o/ S
ninterior.Consider that the Different Individual in similar object has height difference, therefore, a kind of better mode is each layer (H of setting
n min, H
n max) exist overlapping.The more excellent mode of one is setting
Certainly, also can set
Or
Meanwhile, consider that the current hunting zone obtained is only the center point coordinate scope on rectangle base, thus, target scale corresponding for current layer should be taken into account by the hunting zone finally obtained, and makes hunting zone comprise whole target object.Then obtaining effective hunting zone is
Wherein,
For the left side, region of search horizontal ordinate,
For top, region of search ordinate,
For horizontal ordinate on the right of region of search, b
s'=b
sfor the following ordinate in region of search.
4.2.2 by above-mentioned process image, hunting zone R
ninterior image zooming is the 1/S of yardstick on process image
ndoubly, and the sorter of the pre-dimensioning adopting above-mentioned training to obtain, to R
nthe image-region of each position in scope judges, if it is true that sorter exports, then thinks candidate target position, otherwise, not think it is candidate target position.
In order to improve processing speed, can setting search each position time the moving step length in horizontal and vertical direction be greater than 1.
4.2.3 described sorter is exported as genuine position R
o(lo, to, ro, bo), transforms to process image coordinate system, and adds candidate target queue to, carries out merging treatment to the candidate target of close positions.
Step 4.2.3 obtains the position of candidate target in the authentication image of each layer and size, needs to be transformed in process image coordinate system.
According to above-described embodiment, to n-th layer, be R by the candidate target region of sorter
o(lo, to, ro, bo), it is R processing the region in image coordinate system
p o(l
n+ lo/S
n, t
n+ to/S, l
n+ ro/S
n, t
n+ bo/S).
Because progressively each position of traversal search, can obtain multiple candidate target region at close positions place.Therefore, need that aftertreatment is carried out to the result in candidate queue and merge close candidate region.Merging process can with reference to prior art.
Above-mentioned 4.1st step, 4.2.1 step only needs to perform once.
In actual treatment, in order to improve speed further, usually can by other as motion detect, can not there is the scope of object in foreground extraction, and obtain the scope that may there is object in image in the filtering images such as skin analysis.If after above-mentioned process, obtain series of rectangular region, suppose that this rectangular set is altogether containing RM rectangle
then adjust as follows above-mentioned treatment step, to step 4.2.2, setting the hunting zone obtained is the effective hunting zone R of current layer
nwith set
in rectangle R
icommon factor, and carry out subsequent treatment.Thus reduce further the region of search.
Corresponding above method, present invention also offers a kind of embodiment of device, the structural drawing of the device that Fig. 2 provides for the embodiment of the present invention, and as figure, in detected image, the device of object comprises:
Demarcating module 201, for calibrating camera, obtains the corresponding relation between the shooting pixels tall of object target in the shooting image of described video camera and the true altitude of described object target;
Mask images arranges module 202, for: mask images is set, the mask pixels height of described object target in described mask images is set according to described true altitude;
Sorter model module 203, for: the sorter model obtaining described object target according to described mask images;
Search module 204, for: adopt the mode of scaling to obtain the scaling image of the different scale of described shooting image, according to described mask pixels height, described true altitude and described corresponding relation, determine the region of search of described sorter model in described scaling image, and utilize each image-region of described sorter model to described region of search to judge.
Can also comprise: merge module 205, for: the candidate target of current layer is transformed to process image coordinate system, and merges the candidate target region of adjacent position.
Can also comprise: false-alarm processing module 206, for: Gauss model is set according to current progression, utilizes described Gauss model to obtain and detect degree of confidence, get rid of false-alarm according to the pixels tall of described testing result and described detection degree of confidence.
As from the foregoing, the embodiment of the present invention has following advantage:
1) because the embodiment of the present invention have employed the demarcation information of video camera, the pixel size of target on image can be obtained, thus determine the search yardstick of target accordingly, make to need redundancy window to be processed to greatly reduce, not only decrease operand, improve the speed of object detection, decrease interference simultaneously.
2) utilize Gauss model to carry out false-alarm process, decrease false alarm.
3) the present invention is that object detection technology provides possibility in the application of embedded system and SOC hardware system.
The above is only the preferred embodiment of the present invention; it should be pointed out that for those skilled in the art, under the premise without departing from the principles of the invention; can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.
Claims (7)
1. the method for object in detected image, is characterized in that, comprise the steps:
Step one, calibrating camera, obtains the corresponding relation between the shooting pixels tall of object target in the shooting image of described video camera and the true altitude of described object target;
Step 2, arranges mask images, arranges the mask pixels height of described object target in described mask images according to described true altitude;
Step 3, obtains the sorter model of described object target according to described mask images;
Step 4, the mode of scaling is adopted to obtain the scaling image of the different scale of described shooting image, according to described mask pixels height, described true altitude and described corresponding relation, determine the region of search of described sorter model in described scaling image, and utilize described sorter model to search in described region of search; Wherein,
In described step 3, described sorter model is the yardstick by obtaining based on the feature extracting method of window is the image of M × N, and the yardstick of described object target in described sorter model is M' × N';
In described step 4, the scaling number of plies scope of carrying out described scaling is:
wherein,
for the minimum value of the number of plies,
for the maximal value of the number of plies, function f loor (f) is less than the maximum integer of floating number f for getting, and ceiling (f) is greater than the smallest positive integral of floating number f for getting, and scale is scaling factor and constant for being greater than 1, h
minfor the minimum value of described mask pixels height, h
maxfor the maximal value of described mask pixels height;
In described step 4, s=s is respectively to the number of plies
min, s
min+ 1 ... s
maxthe described scaling image of different scale be handled as follows respectively:
To the described scaling image of current layer number, the height hunting zone setting the pixels tall of described object target is:
wherein,
According to described region of search and described height hunting zone, determine effective hunting zone R that current layer number is corresponding
s;
By described hunting zone R
sinterior image zooming is original scale
comparison chart picture doubly, utilizes described sorter model to carry out search comparison in described comparison chart picture, determines that the image-region matched with described sorter model is as testing result;
Described testing result scaling is returned original scale, carries out skew reduction, and merge adjacent testing result.
2. method according to claim 1, is characterized in that, also comprises, setting M=M', N=N'.
3. method according to claim 1, is characterized in that, determines effective hunting zone R that current layer number is corresponding
sprocess specific as follows:
In described region of search, search meets
position and record the minimum value l of the horizontal ordinate of correspondence position
s, horizontal ordinate maximal value r
s, ordinate minimum value be t
s, ordinate maximal value b
s;
Described effective hunting zone is:
wherein, n is current layer number,
for the left side horizontal ordinate of described effective hunting zone,
for the top ordinate of described effective hunting zone,
for the right horizontal ordinate of described effective hunting zone,
for the following ordinate of described effective hunting zone.
4. method according to claim 1, is characterized in that, also comprises, and arranges Gauss model according to current layer number, utilizes described Gauss model to obtain and detects degree of confidence, get rid of false-alarm according to the pixels tall of described testing result and described detection degree of confidence.
5. the device of object in detected image, is characterized in that, comprising:
Demarcating module, for calibrating camera, obtains the corresponding relation between the shooting pixels tall of object target in the shooting image of described video camera and the true altitude of described object target;
Mask images arranges module, for: mask images is set, the mask pixels height of described object target in described mask images is set according to described true altitude;
Sorter model module, for: the sorter model obtaining described object target according to described mask images;
Search module, for: adopt the mode of scaling to obtain the scaling image of the different scale of described shooting image, according to described mask pixels height, described true altitude and described corresponding relation, determine the region of search of described sorter model in described scaling image, and utilize described sorter model to search in described region of search; Wherein,
Described sorter model is the yardstick by obtaining based on the feature extracting method of window is the image of M × N, and the yardstick of described object target in described sorter model is M' × N';
The scaling number of plies scope of carrying out described scaling is:
wherein,
for the minimum value of the number of plies,
for the maximal value of the number of plies, function f loor (f) is less than the maximum integer of floating number f for getting, and ceiling (f) is greater than the smallest positive integral of floating number f for getting, and scale is scaling factor and constant for being greater than 1, h
minfor the minimum value of described mask pixels height, h
maxfor the maximal value of described mask pixels height;
Described search module is respectively s=s to the number of plies
min, s
min+ 1 ... s
maxthe described scaling image of different scale be handled as follows respectively:
To the described scaling image of current layer number, the height hunting zone setting the pixels tall of described object target is:
wherein,
According to described region of search and described height hunting zone, determine effective hunting zone R that current layer number is corresponding
s;
By described hunting zone R
sinterior image zooming is original scale
comparison chart picture doubly, utilizes described sorter model to carry out search comparison in described comparison chart picture, determines that the image-region matched with described sorter model is as testing result.
6. device according to claim 5, is characterized in that, also comprises:
Merge module, for: testing result scaling is returned original scale, carries out skew reduction, and merge adjacent testing result.
7. device according to claim 5, is characterized in that, also comprises:
False-alarm processing module, for: Gauss model is set according to current layer number, utilizes described Gauss model to obtain and detect degree of confidence, get rid of false-alarm according to the pixels tall of described testing result and described detection degree of confidence.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN200910238688.6A CN101719276B (en) | 2009-12-01 | 2009-12-01 | The method and apparatus of object in a kind of detected image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN200910238688.6A CN101719276B (en) | 2009-12-01 | 2009-12-01 | The method and apparatus of object in a kind of detected image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101719276A CN101719276A (en) | 2010-06-02 |
| CN101719276B true CN101719276B (en) | 2015-09-02 |
Family
ID=42433846
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN200910238688.6A Active CN101719276B (en) | 2009-12-01 | 2009-12-01 | The method and apparatus of object in a kind of detected image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101719276B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105303163B (en) * | 2015-09-22 | 2019-03-01 | 深圳市华尊科技股份有限公司 | A kind of method and detection device of target detection |
| CN105956632B (en) * | 2016-05-20 | 2019-06-11 | 浙江宇视科技有限公司 | A kind of method and apparatus detecting target |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7006950B1 (en) * | 2000-06-12 | 2006-02-28 | Siemens Corporate Research, Inc. | Statistical modeling and performance characterization of a real-time dual camera surveillance system |
| CN101196995A (en) * | 2007-12-27 | 2008-06-11 | 北京中星微电子有限公司 | Method for detecting maximum face in image |
| CN101350063A (en) * | 2008-09-03 | 2009-01-21 | 北京中星微电子有限公司 | Method and apparatus for locating human face characteristic point |
| CN101572804A (en) * | 2009-03-30 | 2009-11-04 | 浙江大学 | Multi-camera intelligent control method and device |
-
2009
- 2009-12-01 CN CN200910238688.6A patent/CN101719276B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7006950B1 (en) * | 2000-06-12 | 2006-02-28 | Siemens Corporate Research, Inc. | Statistical modeling and performance characterization of a real-time dual camera surveillance system |
| CN101196995A (en) * | 2007-12-27 | 2008-06-11 | 北京中星微电子有限公司 | Method for detecting maximum face in image |
| CN101350063A (en) * | 2008-09-03 | 2009-01-21 | 北京中星微电子有限公司 | Method and apparatus for locating human face characteristic point |
| CN101572804A (en) * | 2009-03-30 | 2009-11-04 | 浙江大学 | Multi-camera intelligent control method and device |
Non-Patent Citations (1)
| Title |
|---|
| CCD摄像机快速标定技术;周富强等;《光学精密工程》;20000229;第8卷(第1期);96-100 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN101719276A (en) | 2010-06-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106874894B (en) | Human body target detection method based on regional full convolution neural network | |
| CN104376741B (en) | Parking lot state detection method and system | |
| CN101633356B (en) | System and method for detecting pedestrians | |
| CN105260749B (en) | Real-time target detection method based on direction gradient binary pattern and soft cascade SVM | |
| CN105957054B (en) | A kind of image change detection method | |
| CN105740910A (en) | Vehicle object detection method and device | |
| CN103345631B (en) | Image characteristics extraction, training, detection method and module, device, system | |
| CN105654067A (en) | Vehicle detection method and device | |
| US20150332466A1 (en) | Human head detection in depth images | |
| CN103093212A (en) | Method and device for clipping facial images based on face detection and face tracking | |
| CN100561505C (en) | A kind of image detecting method and device | |
| CN112766184B (en) | Remote sensing target detection method based on multi-level feature selection convolutional neural network | |
| CN105893957B (en) | View-based access control model lake surface ship detection recognition and tracking method | |
| CN106940791B (en) | A kind of pedestrian detection method based on low-dimensional histograms of oriented gradients | |
| CN100561501C (en) | A kind of image detecting method and device | |
| CN102750705A (en) | Optical remote sensing image change detection based on image fusion | |
| CN104463240B (en) | A kind of instrument localization method and device | |
| CN102024149B (en) | Method of object detection and training method of classifier in hierarchical object detector | |
| CN105405138A (en) | Water surface target tracking method based on saliency detection | |
| CN103914829B (en) | Method for detecting edge of noisy image | |
| CN116259005A (en) | Intelligent monitoring system based on roof photovoltaic fire control | |
| CN104933719B (en) | One kind integration segment spacing is from detection image edge method | |
| CN109165603A (en) | A kind of Ship Detection and device | |
| CN116824335A (en) | YOLOv5 improved algorithm-based fire disaster early warning method and system | |
| CN116168240A (en) | Arbitrary-direction dense ship target detection method based on attention enhancement |
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 | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20180417 Address after: 100191 Xueyuan Road, Haidian District, Haidian District, Beijing, No. 607, No. six Patentee after: Beijing Vimicro AI Chip Technology Co Ltd Address before: 100083, Haidian District, Xueyuan Road, Beijing No. 35, Nanjing Ning building, 15 Floor Patentee before: Beijing Vimicro Corporation |
|
| TR01 | Transfer of patent right |