CN105260749B - Real-time target detection method based on direction gradient binary pattern and soft cascade SVM - Google Patents
Real-time target detection method based on direction gradient binary pattern and soft cascade SVM Download PDFInfo
- Publication number
- CN105260749B CN105260749B CN201510733481.1A CN201510733481A CN105260749B CN 105260749 B CN105260749 B CN 105260749B CN 201510733481 A CN201510733481 A CN 201510733481A CN 105260749 B CN105260749 B CN 105260749B
- Authority
- CN
- China
- Prior art keywords
- window
- feature
- target
- sample
- svm
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
Abstract
The present invention provides a kind of real-time target detection method being based on direction gradient binary pattern (ORBP) and soft cascade SVM, mainly solves the problems, such as that target detection real-time difference and robustness are low in the prior art.Method and step is:1) direction gradient binary pattern feature description;2) structure of soft cascade grader SVM;3) the feature training of soft cascade grader;4) target window tracking update.ORBP features proposed by the present invention have the advantages that rotation, scale, translation and brightness invariance etc. are a variety of, and soft cascade SVM improves the robustness of target detection under complex scene, and target window tracking improves the real-time of target detection.Method proposed by the present invention can be applied to human-computer interaction and intelligent traffic monitoring field, and target detection performance is excellent.
Description
Technical field
The present invention relates to digital image processing techniques field, it is related to target detection and tracking in computer vision, it can
Applied to fields such as human-computer interaction and intelligent transportation, direction gradient binary pattern and soft cascade are based in particular to one kind
The real-time target detection method of SVM.
Background technology
Target detection is to automatically analyze image by Computerized Information Processing Tech therefrom to detect interested target.Mesh
Important topic of the mark detection as image understanding, is all widely used in military and civilian scene.In reality scene, due to
Change containing other disturbed motion objects, illumination external environment in scene background and target morphology is different and variation is very fast, gives mesh
Mark detection brings many problems, how to realize the target detection of efficient stable, has important real research significance.
Zhang Tianyu proposes a kind of multiscale target detection side in patent " spatiotemporal object moving target detecting method "
Image is carried out piecemeal and realizes object detecting and tracking using optimal difference interval in moving region by method, and this method is in complicated field
Robustness is low under scape, and significant difference decision criteria is difficult to adapt to multiple scenes.Zdenek Kalal, Krystian
Mikolajczyk et al. proposed in " Tracking-Learning-Detection " it is a kind of to video in single target examine
Survey and tracking will be detected using inter-frame information difference and combined with tracking, realize the on-line study to target sample, should
The intermediate value optical flow method that method proposes needs to carry out object initialization, and tracking correction is fixed it is difficult to ensure that synchronous with detector.Yang Yan
Refreshing, the SUSAN that Pu Baoming proposes adaptive threshold in " based on the mobile vehicle detection for improving SUSAN algorithms " detects vehicle
Object boundary method is converted using histogram and is combined extraction target connected domain with Hough transformation, realized to vehicle target and carry on the back
The real-time of the separation of scape, this method is poor and adaptive threshold would become hard to efficiently accomplish Target Segmentation in complex scene.
Invention content
In view of the above shortcomings of the prior art, the present invention is the robust for solving existing object detection method under complex scene
Property low and real-time difference problem, propose a kind of real-time target detection method based on direction gradient binary pattern and soft cascade SVM,
Target detection performance is excellent and is easy to Project Realization.
The above-mentioned purpose of the present invention realizes that dependent claims are to select else or have by the technical characteristic of independent claims
The mode of profit develops the technical characteristic of independent claims.
To achieve the goals above, the present invention provides a kind of real-time based on direction gradient binary pattern and soft cascade SVM
Object detection method, this method are used based on soft cascade SVM SVM using based on direction gradient binary pattern feature
It is described with target signature, positive negative sample is generated using detection image random site when carrying out feature training, finally use shi-
Tomasi Corner Detections extract characteristic point and complete target tracking update.
In some embodiments, detection method includes the following steps for the real-time target:
(1) ORBP feature extractions.Pretreatment operation is carried out to image source sample, utilizes the edges Sobel and local direction ladder
Degree generates ORBP features.
(2) structure of soft cascade grader SVM.Judge whether sample characteristics selection has using cross-correlation characteristic similarity
Effect, according to hk(x) response for calculating all samples finds positive sample boundary and classifies corresponding threshold value, this grade of corresponding threshold value and
Feature will be added to k+1 grades and calculate response hk+1(x);Then window collection to be detected is sequentially sent to soft cascade grader, passed through
Judge current window response to determine whether belonging to target.
(3) training of soft cascade grader.Positive sample negative sample generation is carried out to calibration positive sample target image, to sample
Carry out ORBP feature descriptions;Then starting grader h is trained by SVM0(x), sample image is carried out according to starting grader
Target detection is verified, and negative sample is updated again in next stage SVM cascade classifier training, until completing final cascade sort
Device is trained.
(4) target window tracking update.According to soft cascade SVM train come grader, to image to be detected sequence into
Row target window detects, and the characteristic point of target window is extracted using shi-Tomasi angular-point detection methods, according to Median-Flow
Tracker judges whether current signature point is best tracking point;Then the prediction of next frame target window is calculated by best trace point
Position carries out target discrimination, final output target detection window using the starting grader of cascade classifier.
Wherein, the generation of ORBP features described in step (1) includes the following steps:
(1) image source sample contrast variation is pre-processed, eliminating ambient lighting influences, and pretreatment operation includes
Gaussian smothing filterings and Gamma standardized transformations.
(2) gradient direction is divided into K parts, calculates separately the gradient under each gradient direction in Sobel vertical directions edge
The corresponding gradient-norm length of K Direction interval is put into M sub- block matrix, generates corresponding edge direction gradient map by amplitude.
(3) for each edge direction gradient map carry out it is horizontal divided with vertical direction edge, respectively statistics it is horizontal with it is vertical
Histogram is to accumulative response.For horizontal direction, the accumulative response in upside is F (m1), and the accumulative response in downside is F (m2);For vertical
Direction, the accumulative response in left side are F (m3), and the accumulative response in right side is F (m4), as shown in Fig. 2.
(4) according to horizontal direction, accumulative response generation ORBP compared with the accumulative response magnitude of vertical direction or so is special up and down
Sign, as shown in Fig. 2, the ORBP features of generation contain 4 kinds of bi-level fashions, and when carrying out feature description, histograms of oriented gradients will
It is converted into one kind of corresponding ORBP forms.
Wherein, cross-correlation characteristic similarity determination method detailed process described in step (2) is as follows:
ORBP feature extractions are carried out to image source sample, optionally take one-dimensional characteristic as initial characteristics f0, add other one
Dimensional feature is as second feature f1, calculate feature f0With f1Normalized-cross-correlation function η, the calculating of normalized-cross-correlation function η
Method such as following formula:
Wherein fiIndicate i-th dimension feature vector, cov (fi, fj) indicate feature vector fiWith fjCovariance, var (fi) table
Show feature vector fiVariance.According to cross-correlation coefficient η is calculated, it is effective to judge that current addition is characterized as if η < 0.6,
Otherwise judgement current signature is invalid, need to choose any dimensional feature again.If being further continued for adding any dimensional feature fi+1, need to judge
With current signature vector set { f0, f1…fiCross-correlation coefficient whether meet condition.
Further, to reduce the time complexity of feature description, this method is in step (1) ORBP feature descriptions
Only extract vertical direction edge in the edges Sobel.
Further, in order to improve the completeness of sample selection, this method position in the positive negative sample generating process of step (3)
The selection for setting window needs to meet:Near target neighborhood choose 10 with it under nearest encirclement frame, each scale most
Multiselect takes 4 location windows as positive negative sample.
Advantageous effect:The present invention proposes the real-time target detection method based on direction gradient binary pattern and soft cascade SVM
To solve the problems, such as that the robustness of target detection is low and real-time difference, description is characterized using direction gradient binary pattern, is improved
Robustness of the feature description to complex scene background and illumination variation;It is more with adaptive features select structure soft cascade SVM simultaneously
Grade classification thresholds using random positive and negative sample training grader and are tracked target window, and soft cascade reduces target window
Screening, window tracking improve the stability and real-time of multiframe Sequence Detection.Compared with other similar algorithm of target detection, this hair
The method strong robustness and real-time of bright proposition are good, and target detection performance is excellent.
It should be appreciated that as long as aforementioned concepts and all combinations additionally conceived that describe in greater detail below are at this
Sample design it is not conflicting in the case of can be viewed as the disclosure subject matter a part.In addition, required guarantor
All combinations of the theme of shield are considered as a part for the subject matter of the disclosure.
Can be more fully appreciated from the following description in conjunction with attached drawing present invention teach that foregoing and other aspect, reality
Apply example and feature.The feature and/or advantageous effect of other additional aspects such as illustrative embodiments of the present invention will be below
Description in it is obvious, or by according to present invention teach that specific implementation mode practice in learn.
Description of the drawings
Attached drawing is not intended to drawn to scale.In the accompanying drawings, identical or approximately uniform group each of is shown in each figure
It can be indicated by the same numeral at part.For clarity, in each figure, not each component part is labeled.
Now, by example and the embodiments of various aspects of the invention will be described in reference to the drawings, wherein:
Fig. 1 is the real-time target based on direction gradient binary pattern and soft cascade SVM according to certain embodiments of the invention
The flow chart of detection method.
Fig. 2 is to generate schematic diagram according to the ORBP features of certain embodiments of the invention.
Fig. 3 is the multi-target detection schematic diagram according to certain embodiments of the invention.
Fig. 4 is target detection schematic diagram under the complex scene according to certain embodiments of the invention.
Specific implementation mode
In order to know more about the technology contents of the present invention, spy lifts specific embodiment and institute's accompanying drawings is coordinated to be described as follows.
Various aspects with reference to the accompanying drawings to describe the present invention in the disclosure, shown in the drawings of the embodiment of many explanations.
It is not intended to cover all aspects of the invention for embodiment of the disclosure.It should be appreciated that a variety of designs and reality presented hereinbefore
Those of apply example, and describe in more detail below design and embodiment can in many ways in any one come it is real
It applies, this is because design disclosed in this invention and embodiment are not limited to any embodiment.In addition, disclosed by the invention one
A little aspects can be used alone, or otherwise any appropriately combined be used with disclosed by the invention.
It is according to an embodiment of the invention, real-time based on direction gradient binary pattern and soft cascade SVM in conjunction with shown in Fig. 1
Object detection method includes the following steps:
Direction gradient binary pattern 1. (ORBP) feature description
1.1 pairs of image source f (x, y) sample contrast variation pretreatments, pretreatment operation include Gaussian smothing filterings
And normalized, image normalization operation here is using Gamma standardized transformations:
F (x, y)=ln (f (x, y)+1)
Gradient direction [- pi/2, pi/2] is divided into 9 sections by 1.2, calculates separately each gradient under Sobel vertical edges
Gradient magnitude under directionGenerate corresponding edge direction gradient map.The gradient magnitudeIt is calculated such as with deflection θ
Following formula:
θ=arctan (Gy/Gx)
Wherein Gx, GyRespectively gradients of the image f (x, y) along x and the directions y.
1.3 carry out image to be divided into 100 square blocks, and each cell block is made of 6 × 6 pane locations, respectively will
Each square block edge direction gradient carries out horizontal and vertical direction and divides, and statistics is horizontal to add up response with vertical direction.It is right
In horizontal direction, the accumulative response in upside is F (m1), and the accumulative response in downside is F (m2);For vertical direction, the accumulative response in left side
For F (m3), the accumulative response in right side is F (m4), as shown in Fig. 2.
1.4 respond the generation local binary compared with the accumulative response magnitude of vertical direction or so according to horizontal direction is accumulative up and down
Feature, as shown in Fig. 2, the ORBP features of generation contain 4 kinds of bi-level fashions, and correspondence is translated into when carrying out feature description
One kind of ORBP forms.
2. the structure of soft cascade grader SVM
2.1 cross-correlation characteristic similarities judge, obtain direction gradient binary pattern feature according to step 1, optionally take one-dimensional
Feature is as initial characteristics f0, other one-dimensional characteristics are added as second feature f1, calculate feature f0With f1Normalized crosscorrelation
The computational methods such as following formula of coefficient η, normalized-cross-correlation function η:
Wherein fiIndicate i-th dimension feature vector, cov (fi, fj) indicate feature vector fiWith fjCovariance, var (fi) table
Show feature vector fiVariance.According to cross-correlation coefficient η is calculated, it is effective to judge that current addition is characterized as if η < 0.6,
Otherwise judgement current signature is invalid, need to choose any dimensional feature again.If being further continued for adding any dimensional feature fi+1, need to judge
With current signature vector set { f0, f1…fiCross-correlation coefficient whether meet condition.
The generation of 2.2 cascaded thresholds, the soft cascade grader h of the Linear SVM of construction n dimensionsk(x):
Wherein wiFor the support vector of categorised decision plane i, xiFor corresponding i dimensional features.Cascade classifier is in every level-one spy
Judge whether it is effective according to arest neighbors feature when sign is chosen, then according to hk(x) response for calculating all samples, finds positive sample
The corresponding threshold value of this boundary classification, this grade of corresponding threshold value and feature will be added to k+1 grades and calculate response hk+1(x)。
The judgement of 2.3 cascade sorts.Window collection to be detected is sequentially sent to soft cascade grader, is obtained according to upper step
Threshold value and feature carry out window screening to window to be detected, would consider that it is non-targeted when the response of current window is less than decision-making value,
Then remaining window is subjected to next stage cascade sort judgement.If current window response is higher than current grade grader decision-making value
When, then it is assumed that current window is target.
3. soft cascade grader feature is trained
3.1 positive samples generate, and are scanned to the window of the positive sample target image different scale of calibration, from target neighbour
Domain nearby chooses 10 with it apart from nearest encirclement frame, according to window intersecting area than the positive sample currently chosen of size judgement
Whether this is eligible.It should be noted that at most choosing 4 location windows under each scale as positive sample, to ensure to instruct
Practice the completeness of sample.
3.2 negative samples generate, and while positive sample is chosen, positive sample the window's position is translated up and down, works as window
Intersecting area thinks that the sample that current window generates is negative sample when being less than certain threshold value, each scale equally generates most 4
Location window is as negative sample.
3.3 initial characteristics are trained, and are generated to obtain training sample set according to positive negative sample, are randomly selected positive negative sample each 200
It is a, the lower sample of different scale is uniformly normalized to 60 × 60, direction gradient binary pattern feature description, instruction are carried out to sample
Practise preliminary classification device h0(x)。
3.4 cascade natures are trained, and target detection are carried out to all images according to initial soft cascade, according to positive sample window position
It sets and the training of next stage grader feature is added in flase drop window, after completing image object detection, to newly obtaining positive and negative sample set weight
New training next stage classification hk(x), k ∈ [2 ,+∞), until meet update end condition.
4. target window tracking update
4.1 cascade target window detections.It is right according to the threshold value of trained soft cascade SVM classifier and character pair vector
Image sequence carries out hierarchical classification device judgement, and the final decision judgement output window that cascades is present frame target window.
4.2 extraction target window characteristic points.Upper step obtains cascade classifier output window, utilizes shi-Tomasi angle points
Detection method extracts the characteristic point of target window.It should be noted that for the real-time and accuracy subsequently tracked, angle point inspection
It is excellent to survey worst quality assurance, controls angle point testing number mesh in window and is not more than 20.
4.3 target window tracking points screen, and Median-Flow trackers are utilized according to the tracking point of current n-th frame window
Forward direction tracks the (n+1)th frame, then backward tracing to n-th frame, the variation of calculation window interior angle points amount and back tracking point and former angle point it
Between Euclidean distance, setting decision threshold screens best tracking point.
4.4 target windows more new strategy calculates next frame target window predicted position using best trace point, utilizes cascade
The preliminary classification device of grader carries out target discrimination.If current window is judged as target, it is believed that current tracking is effective, otherwise needs
Sliding window search is carried out according to cascade classifier, re-start target detection.
With reference to concrete scene to the real-time mesh proposed by the present invention based on direction gradient binary pattern and soft cascade SVM
Mark detection method does further details of example test.In hardware platform Inter i5+4G DDR3 RAM, software platform
Implement the example method on OpenCV/C++, 3 Scene multi-target detection of attached drawing is tested, and the present embodiment reaches target detection rate
To 96%, single frames run time is 27ms;Vehicle target detection is tested under complex scene in attached drawing 4, and the present embodiment is to mesh
Mark verification and measurement ratio reaches 95.3%, single frames run time 25ms.
From the above technical scheme, it can be seen that the reality provided by the invention based on direction gradient binary pattern and soft cascade SVM
When object detection method, this method is based on soft cascade SVM SVM, using based on direction gradient binary pattern feature
It is described to target signature, improves robustness of the feature description to complex scene background and illumination variation;Carrying out feature training
Shi Liyong detection image random sites generate positive negative sample, and shi-Tomasi Corner Detections extraction characteristic point is finally used to complete mesh
Mark tracking update, soft cascade reduce target window screening, and window tracking improves the stability and real-time of multiframe Sequence Detection.This
It invents the method strong robustness proposed and real-time is high, target detection performance is excellent.
According to the disclosure, it is also proposed that a kind of detected based on the real-time target of direction gradient binary pattern and soft cascade SVM is filled
It sets, which includes:
For the first module of ORBP feature extractions, which is arranged for pre-processing image source sample
Operation generates ORBP features using the edges Sobel and local direction gradient;
The second module for building soft cascade grader SVM, second module are configured to utilize cross-correlation feature phase
Judge whether sample characteristics selection is effective like degree, according to hk(x) classification of positive sample boundary is found in the response for calculating all samples
Corresponding threshold value, this grade of corresponding threshold value and feature will be added to k+1 grades and calculate response hk+1(x);Then by window to be detected
Collection is sequentially sent to soft cascade grader, by judging current window response to determine whether belonging to target;
Third module for training soft cascade grader, the third module are arranged for demarcating positive sample target
Image carries out positive sample negative sample generation, and ORBP feature descriptions are carried out to sample;Then starting grader h is trained by SVM0
(x), target detection verification is carried out to sample image according to starting grader, negative sample is updated again and arrives next stage SVM cascade
In classifier training, until completing final cascade classifier training;
Newer 4th module is tracked for target window, the 4th module is configured to train according to soft cascade SVM
The grader come, target window detection is carried out to image to be detected sequence, and mesh is extracted using shi-Tomasi angular-point detection methods
The characteristic point for marking window judges whether current signature point is best tracking point according to Median-Flow trackers;Then by most
Good trace point calculates next frame target window predicted position, carries out target discrimination using the starting grader of cascade classifier, most
Output target detection window eventually.
It should be appreciated that the first module, the second module, third module and the 4th module that the present embodiment is proposed, work(
It can, act on and effect is in the real-time target detection method above based on direction gradient binary pattern and soft cascade SVM
It is illustrated, realization method and has been done in the embodiment previously with regard to real-time target detection method exemplary in description
Illustrate, details are not described herein.
Aforementioned embodiments according to the present invention, such as the real-time mesh based on direction gradient binary pattern and soft cascade SVM
Mark detection method and the real-time target detection device based on direction gradient binary pattern and soft cascade SVM, the present invention also propose
It is a kind of for realizing the computer system detected based on the real-time target of direction gradient binary pattern and soft cascade SVM, the calculating
Machine system includes:
Memory;
One or more processors;
One or more modules, the one or more module are stored in the memory and are configured to by described one
A or multiple processors execute, and one or more of modules include following processing modules:
For the first module of ORBP feature extractions, which is arranged for pre-processing image source sample
Operation generates ORBP features using the edges Sobel and local direction gradient;
The second module for building soft cascade grader SVM, second module are configured to utilize cross-correlation feature phase
Judge whether sample characteristics selection is effective like degree, according to hk(x) classification of positive sample boundary is found in the response for calculating all samples
Corresponding threshold value, this grade of corresponding threshold value and feature will be added to k+1 grades and calculate response hk+1(x);Then by window to be detected
Collection is sequentially sent to soft cascade grader, by judging current window response to determine whether belonging to target;
Third module for training soft cascade grader, the third module are arranged for demarcating positive sample target
Image carries out positive sample negative sample generation, and ORBP feature descriptions are carried out to sample;Then starting grader h is trained by SVM0
(x), target detection verification is carried out to sample image according to starting grader, negative sample is updated again and arrives next stage SVM cascade
In classifier training, until completing final cascade classifier training;
Newer 4th module is tracked for target window, the 4th module is configured to train according to soft cascade SVM
The grader come, target window detection is carried out to image to be detected sequence, and mesh is extracted using shi-Tomasi angular-point detection methods
The characteristic point for marking window judges whether current signature point is best tracking point according to Median-Flow trackers;Then by most
Good trace point calculates next frame target window predicted position, carries out target discrimination using the starting grader of cascade classifier, most
Output target detection window eventually.
Although the present invention has been disclosed as a preferred embodiment, however, it is not to limit the invention.Skill belonging to the present invention
Has usually intellectual in art field, without departing from the spirit and scope of the present invention, when can be used for a variety of modifications and variations.Cause
This, the scope of protection of the present invention is defined by those of the claims.
Claims (6)
1. a kind of real-time target detection method based on direction gradient binary pattern and soft cascade SVM, is characterized in that, including following
Step:
(1) ORBP feature extractions:Contrast variation's pretreatment is carried out to image source sample, gradient direction is divided into K parts, respectively
Calculate all directions block gradient figure under the edges image Sobel;Then according to the accumulation of direction block gradient figure level and vertical direction
Response generates ORBP features;
(2) structure of soft cascade grader SVM:According to the ORBP features of image source sample, the response of all samples is calculated, is found
The corresponding threshold value of positive sample boundary classification and feature vector, are then sequentially sent to soft cascade grader by window to be detected, pass through
Current window response magnitude is to determine whether belong to target;
(3) soft cascade grader feature is trained:Positive negative sample generation is carried out to calibration positive sample target image, is randomly selected positive and negative
Sample is each N number of, and ORBP feature descriptions are carried out to sample, is then completed to sample characteristics using the soft cascade SVM classifier of structure
Training;
(4) target window tracking update:According to soft cascade SVM train come grader, to image sequence carry out target window
The characteristic point of target window is extracted in detection using shi-Tomasi angular-point detection methods, is judged according to Median-Flow trackers
Whether current signature point is best tracking point;Then next frame target window predicted position is calculated by best trace point, utilized
The starting grader of cascade classifier carries out target discrimination, final output target detection window.
2. the real-time target detection method based on direction gradient binary pattern and soft cascade SVM as described in claim 1, special
Sign is that the specific method of ORBP feature extractions includes the following steps in the step (1):
(1) to image source sample carry out contrast variation's pretreatment operation, pretreatment operation include Gaussian smothing filterings and
Contrast normalized;
(2) gradient direction is divided into K parts, calculates separately the gradient magnitude under each gradient direction under the edges Sobel, by K
The gradient-norm length of direction scope is put into corresponding M sub- block matrix, generates corresponding edge direction block gradient figure;
(3) it carries out horizontal and vertical direction edge to each edge direction gradient map to divide, counts horizontal and vertical direction respectively
Accumulative response;
(4) according to horizontal direction, accumulative response generates ORBP features compared with the accumulative response magnitude of vertical direction or so up and down.
3. the real-time target detection method based on direction gradient binary pattern and soft cascade SVM as described in claim 1, feature
Be, in the step (2) structure of soft cascade SVM, specific implementation include:
(1) judgement of cross-correlation characteristic similarity is calculated:To image source sample extraction ORBP features, using between cross-correlation calculation feature
Similarity, according to normalized-cross-correlation function judge the sample characteristics choose it is whether effective;
(2) generation of cascaded thresholds:Construct the soft cascade grader h of the Linear SVM of n dimensionsk(x):
Wherein wiFor the support vector of categorised decision plane i, xiFor corresponding i dimensional features;Cascade classifier is selected in every level-one feature
Judge whether it is effective according to cross-correlation feature when taking, then according to hk(x) response for calculating all samples, finds positive sample side
The corresponding threshold value of kingdom, this grade of corresponding threshold value and feature will be added to k+1 grades and calculate response hK+1(x);
(3) judgement of cascade sort:All windows to be detected are sequentially sent to soft cascade grader, the threshold value obtained using cascade
Window screening is carried out to window to be detected with feature, would consider that it is non-targeted when the response of current window is less than decision-making value, then
Remaining window is subjected to next stage cascade classifier judgement:If current window response is higher than current grade grader decision-making value
When, then it is assumed that current window is target.
4. the real-time target detection method based on direction gradient binary pattern and soft cascade SVM as described in claim 1, feature
It is, in the step (3), generates positive negative sample and specifically include:
1) positive sample generates:The window of the positive sample target image different scale of calibration is scanned, it is attached from target neighborhood
It is close to choose 10 with it apart from nearest encirclement frame, according to window intersecting area than size judge that the positive sample currently chosen is
It is no eligible;
2) negative sample generates:While positive sample is chosen, positive sample the window's position is translated up and down, when window intersection
Think that the sample that current window generates is negative sample when product is less than certain threshold value, each scale equally generates most 4 position windows
Mouth is used as negative sample.
5. the real-time target detection method based on direction gradient binary pattern and soft cascade SVM as claimed in claim 3, feature
It is, the cross-correlation characteristic similarity judgement specific method is:
ORBP feature extractions are carried out to image source sample, optionally take one-dimensional characteristic as initial characteristics f0, add other one-dimensional characteristics
As second feature f1, calculate feature f0With f1Normalized-cross-correlation function η, the computational methods of normalized-cross-correlation function η are such as
Following formula:
Wherein, fiIndicate i-th dimension feature vector, cov (fi,fj) indicate feature vector fiWith fjCovariance, var (fi) indicate special
Levy vector fiVariance;According to cross-correlation coefficient η is calculated, if η<η0Then the current addition of judgement is characterized as effective, is otherwise judged
Current signature is invalid, need to choose any dimensional feature again;If being further continued for adding any dimensional feature fI+1, need to judge and current spy
Levy vector set { f0,f1...fiCross-correlation coefficient whether meet condition.
6. a kind of real-time target detection device based on direction gradient binary pattern and soft cascade SVM, which is characterized in that the device
Including:
For the first module of ORBP feature extractions, which is arranged for carrying out pretreatment behaviour to image source sample
Make, ORBP features are generated using the edges Sobel and local direction gradient;
The second module for building soft cascade grader SVM, second module are configured to utilize cross-correlation characteristic similarity
Judge whether sample characteristics selection is effective, according to hk(x) response for calculating all samples is found the classification of positive sample boundary and is corresponded to
Threshold value, this grade of corresponding threshold value and feature will be added to k+1 grades and calculate response hk+1(x);Then by window collection to be detected
It is sequentially sent to soft cascade grader, by judging current window response to determine whether belonging to target;
Third module for training soft cascade grader, the third module are arranged for demarcating positive sample target image
Positive sample negative sample generation is carried out, ORBP feature descriptions are carried out to sample;Then starting grader h is trained by SVM0(x), root
Target detection verification is carried out to sample image according to starting grader, negative sample is updated again and arrives next stage SVM cascade grader
In training, until completing final cascade classifier training;
Track newer 4th module for target window, the 4th module be configured to according to soft cascade SVM train come
Grader, target window detection is carried out to image to be detected sequence, and target window is extracted using shi-Tomasi angular-point detection methods
The characteristic point of mouth judges whether current signature point is best tracking point according to Median-Flow trackers;Then by most preferably with
Track point calculates next frame target window predicted position, and target discrimination is carried out using the starting grader of cascade classifier, final defeated
Go out target detection window;
Wherein it is using the specific method at the edges Sobel and local direction gradient generation ORBP features:Image source sample is carried out
Contrast variation pre-processes, and gradient direction is divided into 9 sections, calculates separately each gradient direction under Sobel vertical edges
Under gradient magnitude, generate corresponding edge direction gradient map, image divided, statistics it is horizontal with vertical direction is accumulative rings
It answers, compares according to response magnitude is horizontally and vertically added up, generate ORBP features.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510733481.1A CN105260749B (en) | 2015-11-02 | 2015-11-02 | Real-time target detection method based on direction gradient binary pattern and soft cascade SVM |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510733481.1A CN105260749B (en) | 2015-11-02 | 2015-11-02 | Real-time target detection method based on direction gradient binary pattern and soft cascade SVM |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105260749A CN105260749A (en) | 2016-01-20 |
CN105260749B true CN105260749B (en) | 2018-11-13 |
Family
ID=55100430
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510733481.1A Active CN105260749B (en) | 2015-11-02 | 2015-11-02 | Real-time target detection method based on direction gradient binary pattern and soft cascade SVM |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105260749B (en) |
Families Citing this family (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106251362B (en) * | 2016-07-15 | 2019-02-01 | 南京莱斯电子设备有限公司 | A kind of sliding window method for tracking target and system based on fast correlation neighborhood characteristics point |
CN106228162B (en) * | 2016-07-22 | 2019-05-17 | 王威 | A kind of quick object identification method of mobile robot based on deep learning |
CN108073940B (en) * | 2016-11-18 | 2021-09-17 | 北京航空航天大学 | Method for detecting 3D target example object in unstructured environment |
CN108229247A (en) * | 2016-12-14 | 2018-06-29 | 贵港市瑞成科技有限公司 | A kind of mobile vehicle detection method |
CN107133558B (en) * | 2017-03-13 | 2020-10-20 | 北京航空航天大学 | Infrared pedestrian significance detection method based on probability propagation |
TWI657378B (en) * | 2017-09-22 | 2019-04-21 | 財團法人資訊工業策進會 | Method and system for object tracking in multiple non-linear distortion lenses |
CN108960320B (en) * | 2018-06-26 | 2021-04-27 | 湖北工业大学 | Real-time detection method for fault image of angle cock of train |
CN110335252B (en) * | 2019-06-04 | 2021-01-19 | 大连理工大学 | Image quality detection method based on background feature point motion analysis |
CN110852323B (en) * | 2019-11-18 | 2022-03-15 | 南京莱斯电子设备有限公司 | Angular point-based aerial target detection method |
CN111681262B (en) * | 2020-05-08 | 2021-09-03 | 南京莱斯电子设备有限公司 | Method for detecting infrared dim target under complex background based on neighborhood gradient |
CN111652292B (en) * | 2020-05-20 | 2022-12-06 | 贵州电网有限责任公司 | Similar object real-time detection method and system based on NCS and MS |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102663409A (en) * | 2012-02-28 | 2012-09-12 | 西安电子科技大学 | Pedestrian tracking method based on HOG-LBP |
CN103077534A (en) * | 2012-12-31 | 2013-05-01 | 南京华图信息技术有限公司 | Space-time multi-scale moving target detection method |
CN103902976A (en) * | 2014-03-31 | 2014-07-02 | 浙江大学 | Pedestrian detection method based on infrared image |
CN104268539A (en) * | 2014-10-17 | 2015-01-07 | 中国科学技术大学 | High-performance human face recognition method and system |
CN104463186A (en) * | 2013-09-16 | 2015-03-25 | 深圳市迈瑞思智能技术有限公司 | Target feature detection method and device |
CN104537360A (en) * | 2015-01-15 | 2015-04-22 | 上海博康智能信息技术有限公司 | Method and system for detecting vehicle violation of not giving way |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR100723406B1 (en) * | 2005-06-20 | 2007-05-30 | 삼성전자주식회사 | Face image verification method and apparatus using LBPLocal Binary Pattern discriminant method |
KR100866792B1 (en) * | 2007-01-10 | 2008-11-04 | 삼성전자주식회사 | Method and apparatus for generating face descriptor using extended Local Binary Pattern, and method and apparatus for recognizing face using it |
-
2015
- 2015-11-02 CN CN201510733481.1A patent/CN105260749B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102663409A (en) * | 2012-02-28 | 2012-09-12 | 西安电子科技大学 | Pedestrian tracking method based on HOG-LBP |
CN103077534A (en) * | 2012-12-31 | 2013-05-01 | 南京华图信息技术有限公司 | Space-time multi-scale moving target detection method |
CN104463186A (en) * | 2013-09-16 | 2015-03-25 | 深圳市迈瑞思智能技术有限公司 | Target feature detection method and device |
CN103902976A (en) * | 2014-03-31 | 2014-07-02 | 浙江大学 | Pedestrian detection method based on infrared image |
CN104268539A (en) * | 2014-10-17 | 2015-01-07 | 中国科学技术大学 | High-performance human face recognition method and system |
CN104537360A (en) * | 2015-01-15 | 2015-04-22 | 上海博康智能信息技术有限公司 | Method and system for detecting vehicle violation of not giving way |
Non-Patent Citations (1)
Title |
---|
基于HOG特征和SVM的前向车辆识别方法;李星等;《计算机科学》;20131130;第40卷(第11A期);329-332 * |
Also Published As
Publication number | Publication date |
---|---|
CN105260749A (en) | 2016-01-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105260749B (en) | Real-time target detection method based on direction gradient binary pattern and soft cascade SVM | |
US10198657B2 (en) | All-weather thermal-image pedestrian detection method | |
CN109977782B (en) | Cross-store operation behavior detection method based on target position information reasoning | |
CN103824070B (en) | A kind of rapid pedestrian detection method based on computer vision | |
CN103886325B (en) | Cyclic matrix video tracking method with partition | |
CN103208008A (en) | Fast adaptation method for traffic video monitoring target detection based on machine vision | |
CN103093212A (en) | Method and device for clipping facial images based on face detection and face tracking | |
CN108197604A (en) | Fast face positioning and tracing method based on embedded device | |
CN108564598B (en) | Improved online Boosting target tracking method | |
CN106778687A (en) | Method for viewing points detecting based on local evaluation and global optimization | |
CN103310194A (en) | Method for detecting head and shoulders of pedestrian in video based on overhead pixel gradient direction | |
CN106778570B (en) | A kind of pedestrian detection and tracking in real time | |
CN112560675B (en) | Bird visual target detection method combining YOLO and rotation-fusion strategy | |
CN104978567A (en) | Vehicle detection method based on scenario classification | |
CN103593679A (en) | Visual human-hand tracking method based on online machine learning | |
CN107103303A (en) | A kind of pedestrian detection method based on GMM backgrounds difference and union feature | |
CN106485273A (en) | A kind of method for detecting human face based on HOG feature and DNN grader | |
CN102592128A (en) | Method and device for detecting and processing dynamic image and display terminal | |
CN108256462A (en) | A kind of demographic method in market monitor video | |
CN109101932A (en) | The deep learning algorithm of multitask and proximity information fusion based on target detection | |
CN103020614A (en) | Human movement identification method based on spatio-temporal interest point detection | |
Zheng et al. | A review of remote sensing image object detection algorithms based on deep learning | |
Su et al. | A new local-main-gradient-orientation HOG and contour differences based algorithm for object classification | |
CN102024149B (en) | Method of object detection and training method of classifier in hierarchical object detector | |
Zhou et al. | Building segmentation from airborne VHR images using Mask R-CNN |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
TR01 | Transfer of patent right |
Effective date of registration: 20181212 Address after: 210000 the 05 building of Tianan Digital City, 36 Yongfeng Avenue, Qinhuai District, Nanjing, Jiangsu. Patentee after: Nanjing Lesi Electronic Equipment Co., Ltd. Address before: No. 1 Yongzhi Road, Qinhuai District, Nanjing City, Jiangsu Province, 210000 Patentee before: China Electronic Technology Corporation (Group) 28 Institute |
|
TR01 | Transfer of patent right |