CN107301431A - A kind of method and device of the non-maxima suppression based on weight - Google Patents
A kind of method and device of the non-maxima suppression based on weight Download PDFInfo
- Publication number
- CN107301431A CN107301431A CN201710517845.1A CN201710517845A CN107301431A CN 107301431 A CN107301431 A CN 107301431A CN 201710517845 A CN201710517845 A CN 201710517845A CN 107301431 A CN107301431 A CN 107301431A
- Authority
- CN
- China
- Prior art keywords
- detection
- confidence point
- confidence
- point
- model
- 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
- 230000001629 suppression Effects 0.000 title claims abstract description 36
- 238000000034 method Methods 0.000 title claims abstract description 21
- 238000001514 detection method Methods 0.000 claims abstract description 329
- 238000005352 clarification Methods 0.000 claims abstract description 7
- 238000007689 inspection Methods 0.000 claims description 15
- 239000000284 extract Substances 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims description 3
- 230000003362 replicative effect Effects 0.000 claims 1
- 230000006870 function Effects 0.000 description 10
- 238000010586 diagram Methods 0.000 description 6
- 230000008878 coupling Effects 0.000 description 5
- 238000010168 coupling process Methods 0.000 description 5
- 238000005859 coupling reaction Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Classifications
-
- 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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2132—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on discrimination criteria, e.g. discriminant analysis
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/98—Detection or correction of errors, e.g. by rescanning the pattern or by human intervention; Evaluation of the quality of the acquired patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Quality & Reliability (AREA)
- Human Computer Interaction (AREA)
- Image Analysis (AREA)
Abstract
The present invention discloses a kind of method of the non-maxima suppression based on weight, comprises the following steps：To detecting that clarification of objective is detected by different detection models, each detection model generates different detection confidence point after detection, and difference examines generation detection redundancy between then model；The linear relationship of detection confidence point is set up, the end value of highest detection confidence point is obtained；According to the end value of calculated highest detection confidence point, highest detection confidence point in detection model is updated；Detection redundancy produced by eliminating, it is determined that detection target location.The invention also discloses a kind of device of the non-maxima suppression based on weight, to realize the above method.Technical solution of the present invention improves the precision and efficiency of pedestrian detection.
Description
Technical field
The present invention relates to pedestrian detection technology field, the method for more particularly to a kind of non-maxima suppression based on weight and
Device.
Background technology
Traditional non-maxima suppression, NMS is the part of many computer vision algorithms makes.Current research work mainly collects
In in directions such as feature extraction, feature learning and graders, and non-maxima suppression direction rarely has improvement.The non-pole commonly used at present
Big value restrainable algorithms are Greedy strategies, single overlapping area information have only been used during suppression, due to the inspection of different detection models
Survey confidence point and there is span difference, therefore different detection models are when detecting same target, even if it is in this position
Detection confidence point be consistent, the level of confidence that it is represented is but completely different.A simple case is lifted, for Grammar moulds
For type, its detection confidence point si values are (- ∞ ,+∞), therefore when si=0, then it represents that half is self-confident, and for inspection
For confidence point value is surveyed for the detection model of (0,1), when si=0.5, just represent that half is self-confident.Therefore, we will not be will
To be readjusted when being combined with detection model, detect that confidence point is normalized to it.And exist during this
Detection redundancy issue of one problem exactly between model A and Model B, carries out overlapping detection to same a group traveling together, and traditional non-
Maximum suppresses that this problem can not be eliminated.
The content of the invention
The main object of the present invention is to propose a kind of method of the non-maxima suppression based on weight, it is intended to improve pedestrian's inspection
The precision of survey.
To achieve the above object, the method for the non-maxima suppression proposed by the present invention based on weight, comprises the following steps：
Each detection model generates different detection confidence after detection is detected by different detection models to detection clarification of objective
Divide, and difference examines generation detection redundancy between then model；The linear relationship of detection confidence point is set up, highest detection confidence point is obtained
End value；According to the end value of calculated highest detection confidence point, highest detection confidence point in detection model is updated；Eliminate
Produced detection redundancy, it is determined that detection target location.
Preferably, the linear relationship for setting up detection confidence point, the step of obtaining the end value of highest detection confidence point
Including：Different detection confidence point are generated according to each detection model after detection, confidence point highest detection model is found；To believe
The heart point highest detection model is core, is set up between the detection model and confidence point highest detection model of relatively low confidence point
Weight relationship；When the value of the weighting function formula is more than default threshold, the detection model of the relatively low confidence point is extracted
Confidence point, and remove the detection model of the relatively low confidence point；Radix is divided into highest detection confidence, with reference to relatively low confidence point
Detection model and confidence point highest detection model between weight relationship, set up the linear relationship of detection confidence point, obtain
The end value of highest detection confidence point.
Preferably, the weighting function between detection model and confidence point highest detection model of the relatively low confidence point
Formula is：
Wherein, the target location that pl goes out for the detection model of relatively low confidence point to detection target detection, ph is confidence point
The target location that highest detection model goes out to detection target detection, area (ph ∩ pl) builds the region of ph and pl common factor stackings
Figure, area (ph ∪ pl) builds the administrative division map of ph and pl unions stacking, and overlaps (ph, pl) is the ph and pl overlapping letter of object
Number.
Preferably, the linear relationship of the detection confidence point is：
Sh+1=Sh+Whl*Sl,
ShIt is highest detection confidence point of all detection models to detection target, Sh+1 is highest detection letter in detection model
The updated value of the heart point, SlIt is detection confidence point of i-th of detection model to detection target, (1, n), what n was represented is inspection to wherein i ∈
Model inspection is surveyed to the number of detection target.
The present invention also proposes a kind of device of the non-maxima suppression based on weight, including detection module, for detection
Clarification of objective is detected that each detection model generates different detection confidence point after detection by different detection models, and
Difference produces detection redundancy between examining then model；Confidence sub-module is detected, the linear relationship for setting up detection confidence point is obtained
The end value of highest detection confidence point；Update module, for the end value according to calculated highest detection confidence point, updates inspection
Survey highest detection confidence point in model；Determining module, for eliminating produced detection redundancy, it is determined that detection target location.
Preferably, the detection confidence sub-module includes：Unit is found, for being generated according to each detection model after detection
Different detection confidence point, finds confidence point highest detection model；Unit is set up, for confidence point highest detection model
For core, the weight relationship set up between the detection model and confidence point highest detection model of relatively low confidence point；Extract single
Member, for when the value of the weighting function formula is more than default threshold, extracting the confidence of the detection model of the relatively low confidence point
Point, and remove the detection model of the relatively low confidence point；Confidence subdivision is detected, for being divided into radix with highest detection confidence,
The weight relationship divided with reference to the detection model and confidence of relatively low confidence point between highest detection model, sets up detection confidence point
Linear relationship, obtain the end value of highest detection confidence point..
Technical solution of the present invention does not delete relative confidence point directly by the non-maxima suppression method based on weight
Low detection model, but the shadow of the low detection model of the confidence point detection model high to confidence point is weighed by weighting function
The degree of sound, the linear relationship that the value of weighting function formula is substituted into detection confidence point obtains final confidence point, improves detection mould
Type judges the precision of detection target location.
Brief description of the drawings
In order to illustrate more clearly about the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing
There is the accompanying drawing used required in technology description to be briefly described, it should be apparent that, drawings in the following description are only this
Some embodiments of invention, for those of ordinary skill in the art, on the premise of not paying creative work, can be with
Structure according to these accompanying drawings obtains other accompanying drawings.
Fig. 1 is the knot method flow diagram of the embodiment of method one of the non-maxima suppression of the invention based on weight；
Fig. 2 is the linear relationship of the present invention for setting up detection confidence point, obtains the final updated of highest detection confidence point
The method flow diagram of the step of value；
Fig. 3 is the functional schematic of the embodiment of pedestrian detection device one of the non-maxima suppression of the invention based on weight；
Fig. 4 refines schematic diagram for the function of detection confidence sub-module of the present invention；
The realization, functional characteristics and advantage of the object of the invention will be described further referring to the drawings in conjunction with the embodiments.
Embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete
Site preparation is described, it is clear that described embodiment is only a part of embodiment of the present invention, rather than whole embodiments.Base
Embodiment in the present invention, those of ordinary skill in the art obtained under the premise of creative work is not made it is all its
His embodiment, belongs to the scope of protection of the invention.
If it is to be appreciated that related in the embodiment of the present invention directionality indicate (such as up, down, left, right, before and after ...),
Then directionality indicate to be only used for explain relative position relation under a certain particular pose (as shown in drawings) between each part,
Motion conditions etc., if the particular pose changes, directionality indicates also correspondingly therewith to change.
If in addition, relating to the description of " first ", " second " etc. in the embodiment of the present invention, being somebody's turn to do " first ", " second " etc.
Description be only used for describing purpose, and it is not intended that indicating or implying its relative importance or implicit indicate indicated skill
The quantity of art feature.Thus, " first " is defined, at least one spy can be expressed or be implicitly included to the feature of " second "
Levy.In addition, the technical scheme between each embodiment can be combined with each other, but must be with those of ordinary skill in the art's energy
Based on enough realizations, when the combination appearance of technical scheme is conflicting or can not realize it will be understood that the knot of this technical scheme
Conjunction is not present, also not within the protection domain of application claims.
The present invention proposes a kind of method of the non-maxima suppression based on weight, is particularly suitable for use in pedestrian detection.
In embodiments of the present invention, as shown in figure 1, a kind of method of the non-maxima suppression based on weight, including it is as follows
Step：
S10 detects that each detection model is generated not after detection by different detection models to detection clarification of objective
With detection confidence point, and different inspections then produce detection redundancy between model；
S20 sets up the linear relationship of detection confidence point, obtains the end value of highest detection confidence point；
S30 updates highest detection confidence point in detection model according to the end value of calculated highest detection confidence point；
S40 eliminates produced detection redundancy, it is determined that detection target location.
Further, the linear relationship for setting up detection confidence point, obtains the step of the end value of highest detection confidence point
Suddenly include：
S201 generates different detection confidence point according to each detection model after detection, finds confidence point highest detection
Model；
S202 sets up the detection model and confidence point of relatively low confidence point using confidence point highest detection model as core
Weight relationship between highest detection model；
S203 extracts the detection model of the relatively low confidence point when the value of the weighting function formula is more than default threshold
Confidence point, and remove the detection model of the relatively low confidence point；
S204 is divided into radix with highest detection confidence, with reference to detection model and the confidence point highest of relatively low confidence point
Weight relationship between detection model, sets up the linear relationship of detection confidence point, obtains the end value of highest detection confidence point.
Further, the weight letter between detection model and confidence point highest detection model of the relatively low confidence point
Numerical expression is：
Wherein, the target location that pl goes out for the detection model of relatively low confidence point to detection target detection, ph is confidence point
The target location that highest detection model goes out to detection target detection, area (ph ∩ pl) builds the region of ph and pl common factor stackings
Figure, area (ph ∪ pl) builds the administrative division map of ph and pl unions stacking, and overlaps (ph, pl) is the ph and pl overlapping letter of object
Number.
Further, the linear relationship of the detection confidence point is：
Sh+1=Sh+Whl*Sl,
ShIt is highest detection confidence point of all detection models to detection target, Sh+1 is highest detection letter in detection model
The updated value of the heart point, SlIt is detection confidence point of i-th of detection model to detection target, (1, n), what n was represented is inspection to wherein i ∈
Model inspection is surveyed to the number of detection target.
The invention also discloses a kind of device of the non-maxima suppression based on weight to realize the above method, including：
Detection module 10, for being detected to detection clarification of objective by different detection models, is each examined after detection
Survey model generates different detection confidence point, and difference examines generation detection redundancy between then model；
Confidence sub-module 20 is detected, the linear relationship for setting up detection confidence point obtains highest detection confidence point most
Final value；
Update module 30, for the end value according to calculated highest detection confidence point, updates highest in detection model
Detect confidence point；
Determining module 40, for eliminating produced detection redundancy, it is determined that detection target location.
Further, the detection confidence sub-module 20 includes：
Unit 201 is found, for generating different detection confidence point according to each detection model after detection, confidence point is found
Highest detection model；
Unit 202 is set up, for using confidence point highest detection model as core, setting up the detection mould of relatively low confidence point
Weight relationship between type and confidence point highest detection model；
Extraction unit 203, for when the value of the weighting function formula is more than default threshold, extracting the relatively low confidence point
Detection model confidence point, and remove the detection model of the relatively low confidence point；
Confidence subdivision 204 is detected, for being divided into radix with highest detection confidence, with reference to the detection mould of relatively low confidence point
Weight relationship between type and confidence point highest detection model, sets up the linear relationship of detection confidence point, obtains highest detection
The end value of confidence point.
In the present invention, in pedestrian's detection technique field, what detection model was generally trained by data sample, but every kind of inspection
All features of human body can not be detected by surveying model, and error occurs unavoidably in detection in all detection models, this
It is that the theory set up by each detection model itself is limited.Therefore, generally will be by different inspections during to pedestrian detection
Survey models coupling to get up, especially for some just complementary models, pass through the combination of detection model, it becomes possible to suppression well
Each of which detection error processed, the combination of a variety of detection models can more accurately detect pedestrian.But different models simultaneously
When to same pedestrian detection, it may appear that pedestrian can be gone out by multiple detection block diagram blocks, here it is our so-called detections
Redundancy.We simultaneously need not these detection redundancies.
Generally, after each detection model is detected to one image of input, { (pi, si) }, wherein i can be expressed as
(1, n), what n was represented is the number for detecting detection target to ∈, and what (pi, si) was represented is i-th of detection target, the target of detection
Position is that pi can be irised out by a detection block diagram, and the object being circled is exactly to detect target, and it is i-th of detection that si, which is represented,
The detection confidence point (confidence score) of target, si is more big, and what is represented is that the position that the detection block diagram is irised out more has
It is probably our pedestrian images really to be detected.There is span not yet with the detection confidence point of different detection models
Together, therefore different detection model is when detecting same target, though its detection confidence in this position divide be it is consistent, it
The level of confidence of representative is but completely different.A simple case is lifted, for Grammar models, its detection confidence point takes
It is worth for (- ∞ ,+∞), therefore when si=0, then it represents that half is self-confident, and for detection that detection confidence point value is (0,1)
For model, when si=0.5, just represent that half is self-confident.Therefore, we will when different detection models are combined
Readjust, detect that confidence point is normalized to it.And had a problem that during this exactly model A and Model B it
Between detection redundancy issue, overlapping detection is carried out to same a group traveling together, and traditional non-maxima suppression can not eliminate this ask
Topic.
Assuming that (pi, si) represents model A detection targets, (pj, sj) represents Model B detection target, and two models exist
There is serious lap in detection process.Usual non-maxima suppression algorithm can compare Si and Sj size, delete confidence
Divide smaller corresponding model, retain the high model of confidence point, and this has also run counter to the original intention of our models couplings, have ignored confidence
Divide influence of the low model to the high model of confidence point, therefore, the present invention proposes a kind of non-maxima suppression algorithm based on weight,
Cleverly solve this problem.
The present invention provides a kind of based on the non-of weight in a variety of detection models based on mall entrance pedestrian detection are combined
Maximum restrainable algorithms (Weighted-NMS), it is low that the non-maxima suppression algorithm based on weight does not have directly deletion confidence point
Detection model, but the influence degree of point low model of the confidence model high to confidence point, confidence point are weighed by weight
It is the expression of an accuracy of the position that detection model judges pedestrian, is the basis for estimation of different model inspection targets, leads to
The linear calculating built up confidence point is crossed, a final highest confidence point is obtained, so that it is low that relative confidence point is embodied in confidence point
The detection model detection model point high to confidence influence.Specific algorithm is described as follows：
Assuming that what (pl, sl) and (ph, sh) represented respectively is the model of relatively low confidence point and the model of highest confidence point,
With (ph, sh) highest confidence sub-model for core, the height of the model and the detection model of highest confidence point of relatively low confidence point
Overlapping, so-called high superposed has standard here, and its standard is, as Whl=overlap (ph, pl)>During T, T is height
One pre-established threshold of overlapping standard, now the model of low confidence point can be deleted, but its confidence point sl will be according to weight
Whl and be added in confidence point Sh, update Sh value：
Wherein, the target location that pl goes out for the detection model of relatively low confidence point to detection target detection, ph is confidence point
The target location that highest detection model goes out to detection target detection, area (ph ∩ pl) builds the region of ph and pl common factor stackings
Figure, area (ph ∪ pl) builds the administrative division map of ph and pl unions stacking, and overlaps (ph, pl) is the ph and pl overlapping letter of object
Number.
Sh<--- -- Sh+Whl*Sl,
ShIt is highest detection confidence point of all detection models to detection target, Sh+1 is highest detection letter in detection model
The updated value of the heart point, SlIt is detection confidence point of i-th of detection model to detection target, (1, n), what n was represented is inspection to wherein i ∈
Model inspection is surveyed to the number of detection target.
, can be dexterously with Jaccard systems and it is overlapping degree between two detection models that wherein weight Whl, which is represented,
Number similarities are weighed, Jaccard coefficients be mainly used in calculating similarity between the individual of symbol measurement or boolean's value metric this
If in we, it can be seen that now Whl be equal to 0, the non-maxima suppression based on weight will degenerate as traditional maximum
Suppress.
The introduction detailed to the present invention more than, it is known that models coupling is in this stream of people of mall entrance pedestrian detection
Necessity under the intensive environment of amount, models coupling can improve the accuracy of detection, next to that introduce based on the non-of weight
Maximum restrainable algorithms can be good at solving the unworthiness that non-maxima suppression eliminates detection redundancy in models coupling, with
The precision and efficiency of pedestrian detection are improved, the non-maxima suppression based on weight can be good at being applied to the detection process
In.It is low that technical solution of the present invention directly deletes relative confidence point by the non-maxima suppression method based on weight, not
Detection model, but the influence journey of the low detection model of the confidence point detection model high to confidence point is weighed by weighting function
Degree, the linear relationship that the value of weighting function formula is substituted into detection confidence point obtains final confidence point, improves detection model and sentences
The precision of disconnected detection target location.
The present invention also proposes a kind of pedestrian detection device of the non-maxima suppression based on weight, to realize above-mentioned one kind
The pedestrian detection method of non-maxima suppression based on weight.The pedestrian detection device of the non-maxima suppression based on weight
Concrete structure is with reference to above-described embodiment, by the pedestrian detection device of this non-maxima suppression based on weight employs above-mentioned institute
There are whole technical schemes of embodiment, therefore all beneficial effects that at least technical scheme with above-described embodiment is brought,
This is no longer going to repeat them.
The preferred embodiments of the present invention are the foregoing is only, are not intended to limit the scope of the invention, it is every at this
Under the inventive concept of invention, the equivalent structure transformation made using description of the invention and accompanying drawing content, or directly/use indirectly
It is included in other related technical fields in the scope of patent protection of the present invention.
Claims (6)
1. a kind of method of the non-maxima suppression based on weight, it is characterised in that comprise the following steps：
S10 detects that each detection model generation is different after detection to detection clarification of objective by different detection models
Detect that confidence point, and difference produce detection redundancy between examining then model；
S20 sets up the linear relationship of detection confidence point, obtains the end value of highest detection confidence point；
S30 updates highest detection confidence point in detection model according to the end value of calculated highest detection confidence point；
S40 eliminates produced detection redundancy, it is determined that detection target location.
2. the method for the non-maxima suppression as claimed in claim 1 based on weight, it is characterised in that the foundation detection letter
The step of linear relationship of the heart point, end value for obtaining highest detection confidence point, includes：
S201 generates different detection confidence point according to each detection model after detection, finds confidence point highest detection model；
S202 sets up the detection model and confidence point highest of relatively low confidence point using confidence point highest detection model as core
Weight relationship between detection model；
S203 extracts the confidence of the detection model of the relatively low confidence point when the value of the weighting function formula is more than default threshold
Point, and remove the detection model of the relatively low confidence point；
S204 is divided into radix with highest detection confidence, with reference to detection model and the confidence point highest detection mould of relatively low confidence point
Weight relationship between type, sets up the linear relationship of detection confidence point, obtains the end value of highest detection confidence point.
3. the method for the non-maxima suppression as claimed in claim 2 based on weight, it is characterised in that the relatively low confidence
Point detection model and confidence point highest detection model between weighting function formula be：
<mrow>
<mi>W</mi>
<mi>h</mi>
<mi>l</mi>
<mo>=</mo>
<mi>o</mi>
<mi>v</mi>
<mi>e</mi>
<mi>r</mi>
<mi>l</mi>
<mi>a</mi>
<mi>p</mi>
<mi>s</mi>
<mrow>
<mo>(</mo>
<mi>p</mi>
<mi>h</mi>
<mo>,</mo>
<mi>p</mi>
<mi>l</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfrac>
<mrow>
<mi>a</mi>
<mi>r</mi>
<mi>e</mi>
<mi>a</mi>
<mrow>
<mo>(</mo>
<mi>p</mi>
<mi>h</mi>
<mo>&cap;</mo>
<mi>p</mi>
<mi>l</mi>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<mi>a</mi>
<mi>r</mi>
<mi>e</mi>
<mi>a</mi>
<mrow>
<mo>(</mo>
<mi>p</mi>
<mi>h</mi>
<mo>&cup;</mo>
<mi>p</mi>
<mi>l</mi>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
<mo>,</mo>
</mrow>
Wherein, the target location that pl goes out for the detection model of relatively low confidence point to detection target detection, ph is confidence point highest
Detection model to the target location that goes out of detection target detection, area (ph ∩ pl) builds ph and pl and occured simultaneously the administrative division map of stacking,
Area (ph ∪ pl) builds the administrative division map of ph and pl unions stacking, and overlaps (ph, pl) is ph and pl object replicative function.
4. the method for the non-maxima suppression as claimed in claim 2 based on weight, it is characterised in that the detection confidence point
Linear relationship be：
Sh+1=Sh+Whl*Sl,
ShIt is highest detection confidence point of all detection models to detection target, Sh+1 is highest detection confidence point in detection model
Updated value, SlIt is detection confidence point of i-th of detection model to detection target, (1, n), what n was represented is detection mould to wherein i ∈
Type detects the number of detection target.
5. a kind of device of the non-maxima suppression based on weight, it is characterised in that including：
Detection module 10, for being detected to detection clarification of objective by different detection models, each detection mould after detection
Type generates different detection confidence point, and difference examines generation detection redundancy between then model；
Confidence sub-module 20 is detected, the linear relationship for setting up detection confidence point obtains the end value of highest detection confidence point；
Update module 30, for the end value according to calculated highest detection confidence point, updates highest detection in detection model
Confidence point；
Determining module 40, for eliminating produced detection redundancy, it is determined that detection target location.
6. the device of the non-maxima suppression as claimed in claim 5 based on weight, it is characterised in that the detection confidence point
Module 20 includes：
Unit 201 is found, for generating different detection confidence point according to each detection model after detection, confidence point highest is found
Detection model；
Set up unit 202, for using confidence point highest detection model as core, set up the detection model of relatively low confidence point with
Weight relationship between confidence point highest detection model；
Extraction unit 203, for when the value of the weighting function formula is more than default threshold, extracting the inspection of the relatively low confidence point
The confidence point of model is surveyed, and removes the detection model of the relatively low confidence point；
Detect confidence subdivision 204, for being divided into radix with highest detection confidence, with reference to relatively low confidence point detection model with
Weight relationship between confidence point highest detection model, sets up the linear relationship of detection confidence point, obtains highest detection confidence
The end value divided.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710517845.1A CN107301431A (en) | 2017-06-29 | 2017-06-29 | A kind of method and device of the non-maxima suppression based on weight |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710517845.1A CN107301431A (en) | 2017-06-29 | 2017-06-29 | A kind of method and device of the non-maxima suppression based on weight |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107301431A true CN107301431A (en) | 2017-10-27 |
Family
ID=60136107
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710517845.1A Pending CN107301431A (en) | 2017-06-29 | 2017-06-29 | A kind of method and device of the non-maxima suppression based on weight |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107301431A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109948480A (en) * | 2019-03-05 | 2019-06-28 | 中国电子科技集团公司第二十八研究所 | A kind of non-maxima suppression method for arbitrary quadrilateral |
| CN112784843A (en) * | 2019-11-07 | 2021-05-11 | 财团法人资讯工业策进会 | Computing device and method for generating object detection model and object detection device |
| US11468594B2 (en) | 2018-08-09 | 2022-10-11 | Boe Technology Group Co., Ltd. | Image processing method, device and apparatus for multi-object detection |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102201112B (en) * | 2010-03-25 | 2013-05-22 | 联咏科技股份有限公司 | Method for scalely removing motion blur of single image |
| CN106372111A (en) * | 2016-08-22 | 2017-02-01 | 中国科学院计算技术研究所 | Local feature point screening method and system |
-
2017
- 2017-06-29 CN CN201710517845.1A patent/CN107301431A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102201112B (en) * | 2010-03-25 | 2013-05-22 | 联咏科技股份有限公司 | Method for scalely removing motion blur of single image |
| CN106372111A (en) * | 2016-08-22 | 2017-02-01 | 中国科学院计算技术研究所 | Local feature point screening method and system |
Non-Patent Citations (2)
| Title |
|---|
| RASMUS ROTHE: "Non-maximum Suppression for Object Detection by Passing Messages Between Windows", 《COMPUTER VISION -- ACCV 2014》 * |
| YUNSHENG JIANG: "Combination Features and Models for Human Detection", 《2015 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11468594B2 (en) | 2018-08-09 | 2022-10-11 | Boe Technology Group Co., Ltd. | Image processing method, device and apparatus for multi-object detection |
| CN109948480A (en) * | 2019-03-05 | 2019-06-28 | 中国电子科技集团公司第二十八研究所 | A kind of non-maxima suppression method for arbitrary quadrilateral |
| CN112784843A (en) * | 2019-11-07 | 2021-05-11 | 财团法人资讯工业策进会 | Computing device and method for generating object detection model and object detection device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2022252398A1 (en) | Ship trajectory feature point extraction-based spatio-temporal dp method | |
| CN107451619A (en) | A kind of small target detecting method that confrontation network is generated based on perception | |
| CN100550054C (en) | A kind of image solid matching method and device thereof | |
| CN111912416B (en) | Method, device and equipment for positioning equipment | |
| CN103218825B (en) | Quick detection method of spatio-temporal interest points with invariable scale | |
| CN107844743A (en) | A kind of image multi-subtitle automatic generation method based on multiple dimensioned layering residual error network | |
| CN105809693A (en) | SAR image registration method based on deep neural networks | |
| CN110222769A (en) | A kind of Further aim detection method based on YOLOV3-tiny | |
| CN102999913A (en) | Local three-dimensional matching method based on credible point spreading | |
| CN106228161A (en) | A kind of pointer-type dial plate automatic reading method | |
| CN107301431A (en) | A kind of method and device of the non-maxima suppression based on weight | |
| CN108197669B (en) | Feature training method and device of convolutional neural network | |
| CN112364931A (en) | Low-sample target detection method based on meta-feature and weight adjustment and network model | |
| EP4105600A2 (en) | Method for automatically producing map data, related apparatus and computer program product | |
| CN105160303A (en) | Fingerprint identification method based on mixed matching | |
| CN105224947A (en) | Sorter training method and system | |
| CN115601318B (en) | Intelligent production method and system for quick-absorption low-reverse-osmosis paper diaper | |
| CN113870254B (en) | Target object detection method and device, electronic equipment and storage medium | |
| CN110245600B (en) | Unmanned aerial vehicle road detection method for self-adaptive initial quick stroke width | |
| CN115272652A (en) | Dense object image detection method based on multiple regression and adaptive focus loss | |
| CN112001217A (en) | Multi-person human body posture estimation algorithm based on deep learning | |
| CN109697727A (en) | Method for tracking target, system and storage medium based on correlation filtering and metric learning | |
| CN109508673A (en) | It is a kind of based on the traffic scene obstacle detection of rodlike pixel and recognition methods | |
| CN108509949A (en) | Object detection method based on attention map | |
| CN104463896A (en) | Image corner point detection method and system based on kernel similar region distribution characteristics |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20171027 |