CN109840918A - A kind of human body detecting method and system - Google Patents
A kind of human body detecting method and system Download PDFInfo
- Publication number
- CN109840918A CN109840918A CN201711204446.6A CN201711204446A CN109840918A CN 109840918 A CN109840918 A CN 109840918A CN 201711204446 A CN201711204446 A CN 201711204446A CN 109840918 A CN109840918 A CN 109840918A
- Authority
- CN
- China
- Prior art keywords
- component
- human body
- detector
- obtains
- subset
- 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
Abstract
The embodiment of the invention discloses a kind of human body detecting method and systems.This method comprises: obtaining multilayer pyramidal layer to the image configuration feature pyramid of input;Human body in first pyramidal layer is detected using combined detector, obtains the first human body candidate region;The second human body candidate region in second pyramidal layer is detected using combined detector, obtains multiple first components;Multiple first component any combination are formed into multiple component subsets;The total detection score for calculating each component subset obtains the first component subset for always detecting highest scoring;Non-maxima suppression processing is carried out to the first component in first component subset, obtains human region;Combined detector includes: single detector, occlusion detector and truncation detector;The resolution ratio of the second pyramidal layer of resolution ratio of first pyramidal layer is low, and the second human body candidate region is obtained by the first human body candidate region.Present invention can ensure that the precision and speed of detection.
Description
Technical field
The present invention relates to Human Detection fields, more particularly to a kind of human body detecting method and system.
Background technique
It accurately, automatically and quickly detects from vehicle monitoring video and positions driver, be presently relevant administrative department's rule
Model driver operation behavior, the urgent need for guaranteeing vehicle driving safety.However, in actual monitor video, due to image
Resolution ratio is lower, and real background is complicated, human body attitude is changeable, the situation blocking and be truncated is more, leads to the inspection to driver
It is difficult to survey exception, it is ineffective.
Summary of the invention
The embodiment of the present invention provides a kind of human body detecting method and system, to solve the prior art to the effect of human testing
Bad problem.
In a first aspect, providing a kind of human body detecting method, comprising: to the image configuration feature pyramid of input, obtain more
Layer pyramidal layer;Human body in first pyramidal layer is detected using combined detector, obtains the first human body candidate region;Using institute
The second human body candidate region in combined detector the second pyramidal layer of detection is stated, multiple first components are obtained;It will be the multiple
First component any combination forms multiple component subsets;The total detection score for calculating each component subset obtains described total
Detect the first component subset of highest scoring;Non-maxima suppression is carried out to the first component in the first component subset
Processing, obtains human region;Wherein, the combined detector includes: single detector, occlusion detector and truncation detector;
The resolution ratio of second pyramidal layer described in the resolution ratio of first pyramidal layer is low, and second human body candidate region is by institute
The first human body candidate region is stated to obtain.
Second aspect provides a kind of detecting system of human body, comprising: constructing module, for the image configuration feature to input
Pyramid obtains multilayer pyramidal layer;First obtains module, for detecting the people in the first pyramidal layer using combined detector
Body obtains the first human body candidate region;Second obtains module, for being detected in the second pyramidal layer using the combined detector
The second human body candidate region, obtain multiple first components;Composite module is used for the multiple first component any combination shape
At multiple component subsets;Computing module obtains described always detecting for calculating total detection score of each component subset
Divide highest first component subset;Processing module, for carrying out non-pole to the first component in the first component subset
Big value inhibition processing, obtains human region;Wherein, the combined detector includes: single detector, occlusion detector and truncation
Detector;The resolution ratio of second pyramidal layer described in the resolution ratio of first pyramidal layer is low, and second human body is candidate
Region is obtained by first human body candidate region.
In this way, by single detector, occlusion detector and truncation detector, can detect in the embodiment of the present invention
Human region that is whole, blocking, be truncated under scene, the first human body by obtaining the restriction of root filter in the first pyramidal layer are waited
Favored area, and obtain in the second pyramidal layer the first component of component filter restriction, thus by human body candidate region into
Row passes through optimal component subset strategy by slightly to smart screening, it is ensured that the precision and speed of detection.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, below by institute in the description to the embodiment of the present invention
Attached drawing to be used is needed to be briefly described, it should be apparent that, the accompanying drawings in the following description is only some implementations of the invention
Example, for those of ordinary skill in the art, without any creative labor, can also be according to these attached drawings
Obtain other attached drawings.
Fig. 1 is the flow chart of the human body detecting method of the embodiment of the present invention;
Fig. 2 is the schematic diagram of the process of the human body detecting method of the embodiment of the present invention;
Fig. 3 is the flow chart of the training step of the human body detecting method of the embodiment of the present invention;
Fig. 4 is the schematic diagram of the process of the training step of the human body detecting method of the embodiment of the present invention;
Fig. 5 is the flow chart of the first human body of acquisition candidate region step of the human body detecting method of the embodiment of the present invention;
Fig. 6 is the flow chart of the multiple first component steps of acquisition of the human body detecting method of the embodiment of the present invention;
Fig. 7 is the first component subset step for obtaining always detecting highest scoring of the human body detecting method of the embodiment of the present invention
Flow chart;
Fig. 8 is that the human body detecting method of the embodiment of the present invention obtains the flow chart of human region step;
Fig. 9 is the structural block diagram of the detecting system of human body of the embodiment of the present invention;
Figure 10 is the testing result of the human body detecting method of the embodiment of the present invention and the human body detecting method of the prior art
Contrast schematic diagram.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are some of the embodiments of the present invention, instead of all the embodiments.Based on this hair
Embodiment in bright, every other implementation obtained by those of ordinary skill in the art without making creative efforts
Example, shall fall within the protection scope of the present invention.
The embodiment of the present invention discloses a kind of human body detecting method.This method mainly utilizes combined detector to realize to input
The quick detection of image.As shown in Figure 1, this method comprises the following steps that
Step S10: to the image configuration feature pyramid of input, multilayer pyramidal layer is obtained.
As shown in Figure 2 A, input picture is indicated.
Step S20: the human body in the first pyramidal layer is detected using combined detector, obtains the first human body candidate region.
Wherein, root filter is that the rectangle of basic covering human region selects frame.As shown in Figure 2 B, it indicates to use joint-detection
Device detects the human body in the first pyramidal layer, obtains the first human body candidate region.
Step S30: the second human body candidate region in the second pyramidal layer is detected using combined detector, obtains multiple the
One component.
Wherein, the resolution ratio of the second pyramidal layer of resolution ratio of the first pyramidal layer is low.Second human body candidate region by
First human body candidate region obtains.As shown in Figure 2 C, it indicates to detect the second people in the second pyramidal layer using combined detector
Body candidate region obtains multiple first components.
Step S40: multiple first component any combination are formed into multiple component subsets.
Step S50: calculating total detection score of each component subset, obtains the first component for always detecting highest scoring
Collection.
By the step, it is usually first for not having negative point in the first component subset of highest scoring that this, which is always detected,
Part can be such that detection total score meets the requirements, and realization fully ensures that Detection accuracy while reaching higher recall rate.
Step S60: non-maxima suppression processing is carried out to the first component in first component subset, obtains human region.
As shown in Figure 2 C, the human region that the rectangle frame in image is.
Specifically, combined detector includes: single detector, occlusion detector and truncation detector.Wherein, single detection
Fully visible people when device is used to detect from the two-dimensional space of the three-dimensional space of real world by human projection to video pictures
Body.Occlusion detector makes the incomplete human region of the human body in video pictures, such as driver for detecting to be blocked by object
It is blocked when being sitting on chair by chair.Truncation detector causes human body incomplete for detecting directly to be truncated by video pictures
Human region, for example, a part of body of driver is located in video pictures, another part body is not in video pictures.Often
All comprising a root filter and several component filters in one detector.In the present embodiment, in each detector
The quantity of component filter is 8.
Above-mentioned single detector, occlusion detector and truncation detector, can first pass through training in advance before testing and obtain.
Specifically, before step S10, as shown in figure 3, this method further includes following training step:
Step S01: randomly selecting plurality of pictures from database, and according to the human body in picture carry out it is complete, block with
And truncation mark, complete human body's data set is constructed respectively, blocks somatic data collection and truncation somatic data collection.
Wherein, the positive sample in complete human body's data set has fully visible human body.Somatic data is blocked to concentrate just
Sample can not complete visible human body with being blocked by object.The positive sample that somatic data is concentrated, which is truncated, to be had by image edge
Truncation and can not complete visible human body.As shown in figure 4, a indicates complete human body's data set, somatic data collection, c are blocked in b expression
Indicate truncation somatic data collection.
Step S02: according to complete human body's data set, according to the single detector of human body rectangle frame length-width ratio training.
Specifically, the step includes following process:
(1) according to the standing of human body and lying posture, complete human body's data set is divided into two the first subsets.
Wherein, the positive sample in first subset is the human body stood, and the positive sample of another the first subset is to present
The human body of lying posture.Specifically, complete human body's data set is divided into two subsets by K-means clustering algorithm by the step.The party
Method is specific as follows:
According to complete human body's data set, the positive sample rectangle frame length-width ratio { r in its data set is calculated1,r2,...rN}。K-
The expression formula of means clustering algorithm are as follows:
By K-means clustering algorithm, calculated result is accordingly divided into m subset, respectively C1... ..., Cm, it is ensured that
Positive sample rectangle frame in each subset has similar length-width ratio.In the present embodiment, m 2, i.e. human body are in standing state
Rectangle frame corresponds to a subset, and the rectangle frame that human body is in lying posture state corresponds to another subset.Then, using positive sample with
M subset is respectively trained in random negative sample, finally obtains m different root filter F1... ..., Fm。
Specifically, by rectangle frame length-width ratio { r1,r2,...rNInput K-means clustering algorithm expression formula.It randomly selects
M cluster subset center of mass point is u1,...,um.Wherein, c(i)For riIt concentrates with m son apart from nearest subset.To each height
Collect j (1≤j≤m), such mass center u is recalculated by following formulaj:
Formula (1) (2) in turn are computed repeatedly, until mass center is constant or changes very little, finally by complete human body's data set point
For two the first subsets.
(2) the single detector of each first subset of training.
The single detector (as shown in figure 4d) of each first subset is trained using positive sample and negative sample.
Step S03: according to somatic data collection is blocked, trained occlusion detector is blocked according to human body.
Specifically, the step includes following process:
(1) somatic data collection will be blocked according to the coverage extent factor and is divided into multiple second subset.
Wherein, the coverage extent factor includes 5%-25%, 25%-55% and 55%-85%, then can will be hidden by the step
Gear somatic data collection is divided into three second subset.
(2) occlusion detector of each second subset of training.
The occlusion detector (as shown in fig 4e) of each second subset is trained using positive sample and negative sample.
Step S04: according to truncation somatic data collection, according to human body truncation degree training truncation detector.
Specifically, the step includes following process:
(1) multiple third subsets are divided into for somatic data collection is truncated according to the truncation degree factor.
Wherein, the truncation degree factor includes 0%-25%, 25%-55% and 55%-85%.
(2) the truncation detector of each third subset of training.
The truncation detector (as shown in fig. 4f) of each third subset is trained using positive sample and negative sample.
Step S05: combined training is carried out to single detector, occlusion detector and truncation detector using the picture chosen
Study, tectonic syntaxis detector.
By above-mentioned training step, combined detector (as shown in figure 4g) can be obtained, for use in the inspection of human region
It surveys.
For the step of detecting, specifically, as shown in figure 5, step S20 includes following process:
Step S201: convolution algorithm is carried out using root filter in the first pyramidal layer, obtains third human body candidate region.
Specific practice is, by sliding window mechanism, by root filter depending on making window, is slided in pyramid feature
It is dynamic, to realize convolution algorithm.The convolution value of third human body candidate region is greater than first threshold.The first threshold can be set in advance
It is fixed.
In general, the first pyramidal layer is the gold removed other than pyramidal 10 layers last (i.e. highest 10 layers of resolution ratio)
Word tower layer.
Root filter is that the rectangle of basic covering human region selects frame.Third human body candidate region is that root filter surrounds
Region.Therefore, the size of third human body candidate region is identical as the size of root filter.
Step S202: non-maxima suppression processing is carried out to third human body candidate region, obtains the first human body candidate region.
Other third human bodies are judged since top score third human body candidate region using non-maxima suppression respectively
Whether the Duplication of candidate region and top score third human body candidate region, which is greater than default Duplication threshold value, (can rule of thumb obtain
Take), give up Duplication and is greater than top score that the third human body candidate region of default Duplication threshold value and marking remains the
Three human body candidate regions.Then from remaining third human body candidate region, select the third human body of wherein top score candidate
Region is further continued for aforesaid operations, and the final third human body candidate region for removing redundancy merges remaining third human body candidate regions
Domain obtains the first human body candidate region.Wherein, which can be obtained by filter convolutional calculation.
Therefore, available first human body candidate region through the above steps, which is one
Roughing region also needs further to carry out the first human region by subsequent step selected.
Specifically, as shown in fig. 6, step S30 includes following process:
Step S301: according to the first human body candidate region, the 4th human body candidate region is obtained in the second pyramidal layer.
Wherein, the corresponding first human body candidate region in the 4th human body candidate region, i.e., the first human body candidate region is according to first
After the ratio variation of the resolution ratio of pyramidal layer and the second pyramidal layer, the region in the second pyramidal layer.Second pyramid
Layer is usually pyramidal 10 layers last (i.e. highest 10 layers of resolution ratio).
Step S302: the range of the 4th human body candidate region is expanded into second threshold, obtains the second human body candidate region.
To pyramidal layer, pyramidal layer can be divided into the identical square unit of multiple sizes.The second threshold one
As be one to two units.Therefore, the second human region is to increase by one to two units around the 4th human body candidate region to obtain
The region arrived.
Step S303: convolution algorithm is carried out using multiple component filters in the second human body candidate region, obtains multiple the
One component.
Specifically, component filter and pyramid feature are carried out convolution algorithm by sliding window mechanism.
Wherein, component filter is that the rectangle of smaller area in covering human region selects frame.Therefore, component filter covers
Human region it is smaller than the human region that root filter covers.The size of each component filter and the size of each first component
It is identical.The quantity of component filter has multiple.In the present embodiment, the quantity of component filter is 8, then the quantity of the first component
It also is 8.
Therefore, through the above steps, the first component can be obtained.
Specifically, as shown in fig. 7, step S40 is specifically included:
Step S401: the coordinate (x of the representative pixel in the first component is obtained0, y0)。
Step S402: according to default score rule, the detection score for representing pixel is calculated, the detection of the first component is obtained
Score.
Wherein, score rule is preset are as follows:
B indicates the deviation of different component filters.
s(pi) indicate i-th of first component detection score.I=1,2 ... ..., n, n indicate the quantity of the first component.
Specifically,
Wherein,Indicate that i-th of size is wi*hiThe first component weight vectors.φ(H,pi) indicate in feature gold
In word tower H, child window size is wi*hi, upper left corner piPyramid feature vector.φd(dx, dy) indicates deformation behaviour.φd
(dx, dy)=(dx, dy, dx2,dy2).Because human body is a non-rigid object, it can be deformed cost, therefore, it is necessary to
Introduce deformation behaviour.
SkIndicate k-th of component subset, | Sk| indicate the number of the first component in k-th of component subset, parameter A and B are logical
Cross the acquisition of sigmoid approximating method.
Step S403: calculating the sum of the detection score for the first component that each component subset includes, and obtains each component
Total detection score of collection.
Therefore, through the above steps, total detection score of first component subset can be obtained.
Specifically, as shown in figure 8, step S60 includes following process:
Step S601: in the first component in first component subset, third member and the 4th component are obtained.
Wherein, third member is the first component of the sub- centralized detecting highest scoring of the first component.4th component is first
The first component in part subset in addition to third member.
Step S602: the Duplication of third member and the 4th component is calculated.
Specifically, the Duplication=two component area intersections/two component area unions.
Step S603: retain the 5th component that Duplication in the 4th component is less than third threshold value.
Wherein, third threshold value can be preset.
Step S604: merge third member and the 5th component, obtain human region.
Through the above steps, human region has been finally obtained.
The embodiment of the invention also discloses a kind of detecting system of human body.Specifically, as shown in figure 9, the system includes:
Constructing module 901 obtains multilayer pyramidal layer for the image configuration feature pyramid to input.
First obtains module 902, for detecting the human body in the first pyramidal layer using combined detector, obtains the first
Body candidate region.
Second obtains module 903, for detecting the second human body candidate regions in the second pyramidal layer using combined detector
Domain obtains multiple first components.
Composite module 904, for multiple first component any combination to be formed multiple component subsets.
Computing module 905 obtains always detecting the first of highest scoring for calculating total detection score of each component subset
Component subset.
Processing module 906 obtains people for carrying out non-maxima suppression processing to the first component in first component subset
Body region.
Wherein, combined detector includes: single detector, occlusion detector and truncation detector;First pyramidal layer
The resolution ratio of the second pyramidal layer of resolution ratio is low, and the second human body candidate region is obtained by the first human body candidate region.
Preferably, the system further include:
First constructing module, for the image configuration feature pyramid to input, the step of obtaining multilayer pyramidal layer it
Before, randomly select plurality of pictures from database, and according to the human body in picture carry out it is complete, block and be truncated mark, divide
Not Gou Zao complete human body's data set, block somatic data collection and truncation somatic data collection.
First training module, for being detected according to the training of human body rectangle frame length-width ratio is single according to complete human body's data set
Device.
Second training module, for blocking trained occlusion detector according to human body according to somatic data collection is blocked.
Third training module, for detector to be truncated according to human body truncation degree training according to truncation somatic data collection.
Second constructing module, for using choose picture to single detector, occlusion detector and truncation detector into
Row combined training study, tectonic syntaxis detector.
Preferably, the first training module includes:
First grouping submodule, for according to human body standing and lying posture, complete human body's data set is divided into two first
Subset.
First training submodule, for training the single detector of each first subset.
Preferably, the second training module includes:
Second packet submodule is divided into multiple second subset for that will block somatic data collection according to the coverage extent factor.
Second training submodule, for training the occlusion detector of each second subset.
Wherein, the coverage extent factor includes 5%-25%, 25%-55% and 55%-85%.
Preferably, third training module includes:
Third is grouped submodule, for being divided into multiple third subsets for somatic data collection is truncated according to the truncation degree factor.
Third trains submodule, for training the truncation detector of each third subset.
Wherein, the truncation degree factor includes 0%-25%, 25%-55% and 55%-85%.
Preferably, the first acquisition module 902 includes:
First operation submodule obtains the third party for carrying out convolution algorithm using root filter in the first pyramidal layer
Body candidate region.
Submodule is handled, for carrying out non-maxima suppression processing to third human body candidate region, obtains the first human body time
Favored area.
Wherein, the convolution value of third human body candidate region is greater than first threshold, the size and root of third human body candidate region
The size of filter is identical.
Preferably, the second acquisition module 903 includes:
First acquisition submodule, for obtaining the 4th human body in the second pyramidal layer according to the first human body candidate region
Candidate region.
Wherein, the corresponding first human body candidate region in the 4th human body candidate region.
Second acquisition submodule obtains the second human body for the range of the 4th human body candidate region to be expanded second threshold
Candidate region.
Second operation submodule, for carrying out convolution algorithm using multiple component filters in the second human body candidate region,
Obtain multiple first components.
Wherein, the size of each component filter is identical as the size of each first component, the people of component filter covering
Body region is smaller than the human region that root filter covers.
Preferably, computing module 905 includes:
Third acquisition submodule, for obtaining the coordinate (x of the representative pixel in the first component0, y0)。
First computational submodule, for calculating the detection score for representing pixel, obtaining first according to default score rule
The detection score of component.
Second computational submodule, the sum of the detection score for calculating the first component that each component subset includes, obtains
Total detection score of each component subset.;
Sorting sub-module obtains first that always detects highest scoring for total detection score sequence to component subset
Part subset.
Wherein, score rule is preset are as follows:
B indicates the deviation of different component filters;s(pi) indicate i-th of first component detection score, i=1,
2 ... ..., n, n indicate the quantity of the first component, It indicates
I-th of size is wi*hiThe first component weight vectors, φ (H, pi) indicate in feature pyramid H, child window size is
wi*hi, upper left corner piPyramid feature vector, φd(dx, dy) indicates deformation behaviour, φd(dx, dy)=(dx, dy, dx2,
dy2);SkIndicate k-th of component subset, | Sk| indicate the number of the first component in k-th of component subset, parameter A and B pass through
Sigmoid approximating method obtains.
Preferably, processing module 906 includes:
4th acquisition submodule, for obtaining third member and the 4th in the first component in first component subset
Part.
Wherein, third member is the first component of the sub- centralized detecting highest scoring of the first component, and the 4th component is first
The first component in part subset in addition to third member;
Third computational submodule, for calculating the Duplication of third member Yu the 4th component;
Retain submodule, the 5th component for being less than third threshold value for retaining Duplication in the 4th component;
Merge submodule and obtains human region for merging third member and the 5th component.
For device embodiment, since it is basically similar to the method embodiment, related so being described relatively simple
Place illustrates referring to the part of embodiment of the method.
With an application examples, technical scheme is described further below.
It is picture to be detected as shown in Figure 10 (a).Certain components (right leg) of human body are blocked in figure.Figure 10 (b) is
The result of human region detection is carried out using traditional DPM (Deformable Part Model) method.Wherein, rectangle frame 1 to 8
What is indicated is the first component.What rectangle frame 0 indicated is the human region that detection obtains.This method still nationwide examination for graduation qualification when detecting the human body
Worry is blocked component score (rectangle frame 5), and is added in whole human testing score.Due to being blocked, component score is negative
Number, and then the total score of detection part is reduced, cause the human body to be missed or increase other positions erroneous detection, seriously affects final
Testing result.As Figure 10 (c) show the result of the method progress human region detection using the embodiment of the present invention.Wherein, square
What shape frame 1 to 8 indicated is the first component.What rectangle frame 0 indicated is the human region that detection obtains.Its mid-score be positive first
Component corresponds to rectangle frame 1 to 3, rectangle frame 5 to 6 and rectangle frame 8, and the first component that score is negative corresponds to rectangle frame 4 and rectangle frame
7.The detection score of rectangle frame 4 and rectangle frame 7 is removed, so that total detection of rectangle frame 1 to 3, rectangle frame 5 to 6 and rectangle frame 8
Score is higher, avoids missing inspection.
To sum up, the embodiment of the present invention is detected especially suitable for the human body target under low resolution complex background.The present invention
Embodiment had not only made full use of traditional single detector, but also can merge the fresh target detector for special scenes, that is, blocked inspection
Survey device and truncation detector, can be widely applied to fixed, narrow space target detection, for example, be applied to such as coach,
The scenes such as bus, taxi, have very strong universality, well compatibility and scalability, it can be achieved that it is more general and
Accurate object detection task.The embodiment of the present invention uses optimal component subset strategy, by selecting suitable component subset meter
It calculates and more accurately detects score.It is used in the embodiment of the present invention from slightly to the strategy of essence, acceleration detection process substantially meets reality
When detect.
It above to technical solution provided by the present invention, is described in detail, specific case used herein is to this
The principle and embodiment of invention is expounded, method of the invention that the above embodiments are only used to help understand and
Its core concept;At the same time, for those skilled in the art in specific embodiment and is answered according to the thought of the present invention
With in range, there will be changes, in conclusion the contents of this specification are not to be construed as limiting the invention.
Claims (10)
1. a kind of human body detecting method characterized by comprising
To the image configuration feature pyramid of input, multilayer pyramidal layer is obtained;
Human body in first pyramidal layer is detected using combined detector, obtains the first human body candidate region;
The second human body candidate region in second pyramidal layer is detected using the combined detector, obtains multiple first components;
The multiple first component any combination is formed into multiple component subsets;
The total detection score for calculating each component subset obtains the first component subset of total detection highest scoring;
Non-maxima suppression processing is carried out to the first component in the first component subset, obtains human region;
Wherein, the combined detector includes: single detector, occlusion detector and truncation detector;First pyramid
The resolution ratio of second pyramidal layer described in the resolution ratio of layer is low, and second human body candidate region is candidate by first human body
Region obtains.
2. the method according to claim 1, wherein the image configuration feature pyramid of described pair of input, obtains
Before the step of multilayer pyramidal layer, the method also includes:
Randomly select plurality of pictures from database, and according to the human body in picture carry out it is complete, block and be truncated mark, divide
Not Gou Zao complete human body's data set, block somatic data collection and truncation somatic data collection;
According to complete human body's data set, according to the human body rectangle frame length-width ratio training single detector;
Somatic data collection is blocked according to described, blocks the training occlusion detector according to human body;
According to the truncation somatic data collection, according to the human body truncation degree training truncation detector;
The single detector, the occlusion detector and the truncation detector are mixed using the picture of selection
Training study, constructs the combined detector.
3. according to the method described in claim 2, it is characterized in that, described according to complete human body's data set, according to human body
The step of rectangle frame length-width ratio training single detector, comprising:
According to the standing of human body and lying posture, complete human body's data set is divided into two the first subsets;
The single detector of each first subset of training.
4. according to the method described in claim 2, it is characterized in that, described block somatic data collection according to, according to human body
The step of blocking the training occlusion detector, comprising:
The somatic data collection that blocks is divided into multiple second subset according to the coverage extent factor;
The occlusion detector of each second subset of training;
Wherein, the coverage extent factor includes 5%-25%, 25%-55% and 55%-85%.
5. according to the method described in claim 2, it is characterized in that, described according to the truncation somatic data collection, according to human body
The step of truncation degree training truncation detector, comprising:
The truncation somatic data collection is divided into multiple third subsets according to the truncation degree factor;
The truncation detector of each third subset of training;
Wherein, the truncation degree factor includes 0%-25%, 25%-55% and 55%-85%.
6. the method according to claim 1, wherein described detected in the first pyramidal layer using combined detector
Human body, obtain the first human body candidate region the step of, comprising:
Convolution algorithm is carried out using root filter in first pyramidal layer, obtains third human body candidate region;
Non-maxima suppression processing is carried out to third human body candidate region, obtains first human body candidate region;
Wherein, the convolution value of third human body candidate region is greater than first threshold, the size of third human body candidate region
It is identical as the size of described filter.
7. according to the method described in claim 6, it is characterized in that, described detect the second pyramid using the combined detector
The second human body candidate region in layer, the step of obtaining multiple first components, comprising:
According to first human body candidate region, the 4th human body candidate region is obtained in second pyramidal layer, wherein institute
It states the 4th human body candidate region and corresponds to first human body candidate region;
The range of the 4th human body candidate region is expanded into second threshold, obtains second human body candidate region;
Convolution algorithm is carried out using multiple component filters in second human body candidate region, obtains multiple described first
Part;
Wherein, the size of each component filter is identical as the size of each first component, the people of the component filter covering
The human region that than described filter of body region covers is small.
8. the method according to claim 1, wherein described calculate always detecting for each component subset
The step of dividing, obtaining the first component subset of total detection highest scoring, comprising:
Obtain the coordinate (x of the representative pixel in the first component0, y0);
According to default score rule, the detection score for representing pixel is calculated, the detection score of the first component is obtained;
The sum for calculating the detection score for the first component that each component subset includes, obtains each component subset
Total detection score;
Total detection score sequence to the component subset obtains the first component subset of total detection highest scoring;
Wherein, the default score rule are as follows:
B indicates the deviation of different component filters;s(pi) indicate i-th of first component detection score, i=1,2 ... ...,
N, n indicate the quantity of the first component, Indicate i-th big
Small is wi*hiThe first component weight vectors, φ (H, pi) indicate in feature pyramid H, child window size is wi*hi, left
Upper angle is piPyramid feature vector, φd(dx, dy) indicates deformation behaviour, φd(dx, dy)=(dx, dy, dx2,dy2);Sk
Indicate k-th of component subset, | Sk| indicate the number of the first component in k-th of component subset, parameter A and B pass through sigmoid
Approximating method obtains.
9. the method according to claim 1, wherein described first in the first component subset
The step of part carries out non-maxima suppression processing, obtains human region, comprising:
In the first component in the first component subset, third member and the 4th component are obtained, wherein the third
Component is the first component of the sub- centralized detecting highest scoring of the first component, and the 4th component is the first component subset
In the first component in addition to the third member;
Calculate the Duplication of the third member Yu the 4th component;
Retain the 5th component that Duplication in the 4th component is less than third threshold value;
Merge the third member and the 5th component, obtains human region.
10. a kind of detecting system of human body characterized by comprising
Constructing module obtains multilayer pyramidal layer for the image configuration feature pyramid to input;
First obtains module, and for detecting the human body in the first pyramidal layer using combined detector, it is candidate to obtain the first human body
Region;
Second obtains module, for detecting the second human body candidate region in the second pyramidal layer using the combined detector,
Obtain multiple first components;
Composite module, for the multiple first component any combination to be formed multiple component subsets;
Computing module obtains the of total detection highest scoring for calculating total detection score of each component subset
One component subset;
Processing module is obtained for carrying out non-maxima suppression processing to the first component in the first component subset
Human region;
Wherein, the combined detector includes: single detector, occlusion detector and truncation detector;First pyramid
The resolution ratio of second pyramidal layer described in the resolution ratio of layer is low, and second human body candidate region is candidate by first human body
Region obtains.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711204446.6A CN109840918A (en) | 2017-11-27 | 2017-11-27 | A kind of human body detecting method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711204446.6A CN109840918A (en) | 2017-11-27 | 2017-11-27 | A kind of human body detecting method and system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109840918A true CN109840918A (en) | 2019-06-04 |
Family
ID=66880119
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711204446.6A Pending CN109840918A (en) | 2017-11-27 | 2017-11-27 | A kind of human body detecting method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109840918A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112446246A (en) * | 2019-08-30 | 2021-03-05 | 初速度(苏州)科技有限公司 | Image occlusion detection method and vehicle-mounted terminal |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106407943A (en) * | 2016-09-28 | 2017-02-15 | 天津工业大学 | Pyramid layer positioning based quick DPM pedestrian detection method |
-
2017
- 2017-11-27 CN CN201711204446.6A patent/CN109840918A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106407943A (en) * | 2016-09-28 | 2017-02-15 | 天津工业大学 | Pyramid layer positioning based quick DPM pedestrian detection method |
Non-Patent Citations (1)
Title |
---|
方豪: "复杂环境下的列车驾驶员目标检测", 《中国优秀博硕士学位论文全文数据库(硕士) 工程科技Ⅱ辑》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112446246A (en) * | 2019-08-30 | 2021-03-05 | 初速度(苏州)科技有限公司 | Image occlusion detection method and vehicle-mounted terminal |
CN112446246B (en) * | 2019-08-30 | 2022-06-21 | 魔门塔(苏州)科技有限公司 | Image occlusion detection method and vehicle-mounted terminal |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111750820B (en) | Image positioning method and system | |
CN103578133B (en) | A kind of method and apparatus that two-dimensional image information is carried out to three-dimensional reconstruction | |
AU2008322565B9 (en) | Method and apparatus of taking aerial surveys | |
CN107993263B (en) | Automatic calibration method for panoramic system, automobile, calibration device and storage medium | |
CN105608421B (en) | A kind of recognition methods of human action and device | |
CN109801374B (en) | Method, medium, and system for reconstructing three-dimensional model through multi-angle image set | |
US20150379768A1 (en) | Estimation of object properties in 3d world | |
CN104537705B (en) | Mobile platform three dimensional biological molecular display system and method based on augmented reality | |
CN106910242A (en) | The method and system of indoor full scene three-dimensional reconstruction are carried out based on depth camera | |
CN110009561A (en) | A kind of monitor video target is mapped to the method and system of three-dimensional geographical model of place | |
CN105701828B (en) | A kind of image processing method and device | |
CN106204605B (en) | A kind of localization method and device | |
CN109035327B (en) | Panoramic camera attitude estimation method based on deep learning | |
CN108074237B (en) | Image definition detection method and device, storage medium and electronic equipment | |
CN106485207A (en) | A kind of Fingertip Detection based on binocular vision image and system | |
CN110473216A (en) | The detection method and device of object in a kind of image | |
CN1897715A (en) | Three-dimensional vision semi-matter simulating system and method | |
CN105913488A (en) | Three-dimensional-mapping-table-based three-dimensional point cloud rapid reconstruction method | |
CN110324534B (en) | Image processing method and device and electronic equipment | |
CN106778660B (en) | A kind of human face posture bearing calibration and device | |
CN107403451A (en) | Adaptive binary feature monocular vision odometer method and computer, robot | |
CN108921889A (en) | A kind of indoor 3-D positioning method based on Augmented Reality application | |
CN114926747A (en) | Remote sensing image directional target detection method based on multi-feature aggregation and interaction | |
CN110096993A (en) | The object detection apparatus and method of binocular stereo vision | |
CN106169076A (en) | A kind of angle license plate image storehouse based on perspective transform building method |
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: 20190604 |