CN105260749A - Real-time target detection method based on oriented gradient two-value mode and soft cascade SVM - Google Patents
Real-time target detection method based on oriented gradient two-value mode and soft cascade SVM Download PDFInfo
- Publication number
- CN105260749A CN105260749A CN201510733481.1A CN201510733481A CN105260749A CN 105260749 A CN105260749 A CN 105260749A CN 201510733481 A CN201510733481 A CN 201510733481A CN 105260749 A CN105260749 A CN 105260749A
- Authority
- CN
- China
- Prior art keywords
- window
- sample
- feature
- target
- 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.)
- Granted
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
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a real-time target detection method based on the oriented gradient two-value mode (ORBP) and soft cascade SVM, and aims at solving the problems that the target detection in the prior art is low in instantaneity and robustness. The method comprises the steps that 1) characteristic of the oriented gradient two-value mode is described; 2) a soft cascade classifier SVM is established; 3) characteristic training is carried out on the soft cascade classifier SVM; and 4) a target window is tracked and updated. ORBP has the advantages that rotation, scale, translation and brightness are not changed, the soft cascade SVM improves the target detection robustness in complex scenes, and tracking of the target window improves the instantaneity of target detection. The method provided by the invention can be applied to man-machine interaction and intelligent traffic monitoring fields, and the target detection performance is excellent.
Description
Technical field
The present invention relates to digital image processing techniques field, relate to target detection and tracking in computer vision, can be applicable to the field such as man-machine interaction and intelligent transportation, in particular to a kind of real-time target detection method based on direction gradient binary pattern and soft cascade SVM.
Background technology
Target detection therefrom detects interested target by Computerized Information Processing Tech automatic analysis image.Target detection, as the important topic of image understanding, is all widely used in military affairs with civilian scene.In reality scene, due in scene background containing other disturbed motion objects, the change of illumination external environment condition and target morphology is different and change is very fast, bring many difficult problems to target detection, how to realize the target detection of efficient stable, there is important real Research Significance.
Zhang Tianyu proposes a kind of multiscale target detection method in patent " spatiotemporal object moving target detecting method ", image is carried out piecemeal and utilize optimum difference interval realize target detection and tracking in moving region, the method robustness under complex scene is low, and significant difference decision criteria is difficult to adapt to multiple scene.ZdenekKalal, the people such as KrystianMikolajczyk propose in " Tracking-Learning-Detection " a kind of to video in single target detection and tracking method, inter-frame information difference is utilized to be combined by detection and tracking, realize the on-line study to target sample, the intermediate value optical flow method that the method proposes needs to carry out object initialization, and the fixing very difficult guarantee of tracking correction is synchronous with detecting device.Yang Yanshuan, the SUSAN that Pu Baoming proposes adaptive threshold in " detecting based on the moving vehicle improving SUSAN algorithm " detects vehicle target boundary method, utilize histogram transformation to be combined with Hough transformation and extract target connected domain, realize being separated vehicle target and background, the real-time of the method poor and in complex scene adaptive threshold effectively complete Target Segmentation by being difficult to.
Summary of the invention
For above-mentioned the deficiencies in the prior art, the present invention solves the low and poor real problem of the robustness of existing object detection method under complex scene, 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 be easy to Project Realization.
Above-mentioned purpose of the present invention is realized by the technical characteristic of independent claims, and dependent claims develops the technical characteristic of independent claims with alternative or favourable mode.
To achieve these goals, the invention provides a kind of real-time target detection method based on direction gradient binary pattern and soft cascade SVM, the method is based on soft cascade Support Vector Machine SVM, adopt and describe in order to target signature based on direction gradient binary pattern feature, utilizing when carrying out features training detected image random site to generate positive negative sample, finally adopting shi-Tomasi Corner Detection extract minutiae to complete target tracking and upgrading.
In certain embodiments, this real-time target detection method comprises the following steps:
(1) ORBP feature extraction.Pretreatment operation is carried out to image source sample, utilizes Sobel edge and local direction gradients setup ORBP feature.
(2) structure of soft cascade sorter SVM.Whether this sample characteristics is chosen effective, according to h to utilize cross-correlation characteristic similarity to judge
kx () calculates the response of all samples, find positive sample boundary to classify corresponding threshold value, and this grade of corresponding threshold value and feature will join k+1 level calculated response h
k+1(x); Then window collection to be detected is sent into soft cascade sorter successively, by judging that current window response judges whether to belong to target.
(3) training of soft cascade sorter.Carry out positive sample negative sample to the positive sample object image of demarcation to generate, ORBP feature interpretation is carried out to sample; Then initial sorter h is trained by SVM
0x (), carries out target detection checking according to initial sorter to sample image, be again updated to by negative sample in the training of next stage SVM cascade classifier, until complete the training of final cascade classifier.
(4) target window is followed the trail of and is upgraded.According to the sorter that soft cascade SVM trains out, treat detected image sequence and carry out target window detection, utilize shi-Tomasi angular-point detection method to extract the unique point of target window, judge current signature point whether as best tracking point according to Median-Flow tracker; Then calculate next frame target window predicted position by best trace point, utilize the initial sorter of cascade classifier to carry out target discrimination, final output target detection window.
Wherein, ORBP feature described in step (1) generates and comprises the following steps:
(1) to image source sample contrast variation pre-service, eliminate ambient lighting impact, pretreatment operation comprises Gaussian smothing filtering and Gamma standardized transformation.
(2) gradient direction is divided into K part, calculates the gradient magnitude under Sobel vertical direction edge each gradient direction respectively, put into the individual sub-block matrix of M by long for gradient-norm corresponding for K Direction interval, generate corresponding edge direction gradient map.
(3) carry out level for each edge direction gradient map and vertical direction edge divides, level of adding up respectively adds up to respond with vertical direction.For horizontal direction, the accumulative response in upside is m
1, the accumulative response in downside is m
2; For vertical direction, the accumulative response in left side is m
3, the accumulative response in right side is m
4, as shown in Figure 2.
(4) according to horizontal direction up and down accumulative response with add up response magnitude about vertical direction and compare and generate ORBP feature, as shown in Figure 2, the ORBP feature generated contains 4 kinds of bi-level fashion, and when carrying out feature interpretation, histograms of oriented gradients will be converted into the one of corresponding ORBP form.
Wherein, the detailed process of cross-correlation characteristic similarity decision method described in step (2) is as follows:
ORBP feature extraction is carried out to image source sample, optionally gets one-dimensional characteristic as initial characteristics f
0, add other one-dimensional characteristics as second feature f
1, calculate feature f
0with f
1normalized-cross-correlation function η, the computing method of normalized-cross-correlation function η as shown in the formula:
Wherein f
irepresent the i-th dimensional feature vector, cov (f
i, f
j) representation feature vector f
iwith f
jcovariance, var (f
i) representation feature vector f
ivariance.According to calculating cross-correlation coefficient η, if η is <0.6, judge that current interpolation is characterized as effectively, otherwise judge that current signature is as invalid, need again choose arbitrary dimensional feature.Arbitrary dimensional feature f is added if continue again
i+1, need judge and current signature vector set { f
0, f
1... f
icross-correlation coefficient whether satisfy condition.
Further, for reducing the time complexity of feature interpretation, the method only extracts vertical direction edge in Sobel edge in step (1) ORBP feature interpretation.
Further, in order to improve the completeness that sample is chosen, the method in the positive negative sample generative process of step (3) location window choose demand fulfillment: near target neighborhood, choose 10 encirclement frames nearest with it, choose at most 4 location window under often kind of yardstick as positive negative sample.
Beneficial effect: the real-time target detection method that the present invention proposes based on direction gradient binary pattern and soft cascade SVM is low with poor real problem with the robustness solving target detection, adopt direction gradient binary pattern to be Feature Descriptor, improve the robustness of feature interpretation to complex scene background and illumination variation; Build soft cascade SVM multiclass classification threshold value with adaptive features select simultaneously, utilize random positive and negative sample training sorter and follow the trail of target window, soft cascade reduces target window screening, and window follows the trail of the stability and real-time that improve multiframe Sequence Detection.Compared with algorithm of target detection similar with other, method strong robustness and the real-time of the present invention's proposition are good, and target detection performance is excellent.
As long as should be appreciated that aforementioned concepts and all combinations of extra design described in further detail below can be regarded as a part for subject matter of the present disclosure when such design is not conflicting.In addition, all combinations of theme required for protection are all regarded as a part for subject matter of the present disclosure.
The foregoing and other aspect of the present invention's instruction, embodiment and feature can be understood by reference to the accompanying drawings from the following description more all sidedly.Feature and/or the beneficial effect of other additional aspect of the present invention such as illustrative embodiments will be obvious in the following description, or by learning in the practice of the embodiment according to the present invention's instruction.
Accompanying drawing explanation
Accompanying drawing is not intended to draw in proportion.In the accompanying drawings, each identical or approximately uniform ingredient illustrated in each figure can represent with identical label.For clarity, in each figure, not each ingredient is all labeled.Now, the embodiment of various aspects of the present invention also will be described with reference to accompanying drawing by example, wherein:
Fig. 1 is the process flow diagram of the real-time target detection method based on direction gradient binary pattern and soft cascade SVM according to certain embodiments of the invention.
Fig. 2 generates schematic diagram according to the ORBP feature 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.
Embodiment
In order to more understand technology contents of the present invention, institute's accompanying drawings is coordinated to be described as follows especially exemplified by specific embodiment.
Each side with reference to the accompanying drawings to describe the present invention in the disclosure, shown in the drawings of the embodiment of many explanations.Embodiment of the present disclosure must not be intended to comprise all aspects of the present invention.Be to be understood that, multiple design presented hereinbefore and embodiment, and describe in more detail below those design and embodiment can in many ways in any one is implemented, this is because design disclosed in this invention and embodiment are not limited to any embodiment.In addition, aspects more disclosed by the invention can be used alone, or otherwisely anyly appropriately combinedly to use with disclosed by the invention.
Shown in composition graphs 1, according to embodiments of the invention, the real-time target detection method based on direction gradient binary pattern and soft cascade SVM comprises the following steps:
1. direction gradient binary pattern (ORBP) feature interpretation
1.1 couples of image source f (x, y) sample contrast variations pre-service, pretreatment operation comprises Gaussian smothing filtering and normalized, and what image normalization operation here adopted is Gamma standardized transformation:
f(x,y)＝ln(f(x,y)+1)
Gradient direction [-pi/2, pi/2] is divided into 9 intervals by 1.2, the gradient magnitude respectively under the calculating vertical edge of Sobel under each gradient direction | and ▽ f|, generates corresponding edge direction gradient map.Described gradient magnitude | ▽ f| and deflection θ is calculated as follows formula:
θ＝arctan(G
y/G
x)
Wherein G
x, G
ybe respectively the gradient of image f (x, y) along x and y direction.
Image carries out being divided into 100 square block by 1.3, and each cell block is made up of 6 × 6 pane location, respectively each square block edge direction gradient is carried out level and vertical direction divides, and statistics level and vertical direction add up to respond.For horizontal direction, the accumulative response in upside is m
1, the accumulative response in downside is m
2; For vertical direction, the accumulative response in left side is m
3, the accumulative response in right side is m
4, as shown in Figure 2.
1.4 according to horizontal direction up and down accumulative response with add up response magnitude about vertical direction and compare and generate local binary feature, as shown in Figure 2, the ORBP feature generated contains 4 kinds of bi-level fashion, will be converted into the one of corresponding ORBP form when carrying out feature interpretation.
2. the structure of soft cascade sorter SVM
2.1 cross-correlation characteristic similarities judge, obtain direction gradient binary pattern feature, optionally get one-dimensional characteristic as initial characteristics f according to step 1
0, add other one-dimensional characteristics as second feature f
1, calculate feature f
0with f
1normalized-cross-correlation function η, the computing method of normalized-cross-correlation function η as shown in the formula:
Wherein f
irepresent the i-th dimensional feature vector, cov (f
i, f
j) representation feature vector f
iwith f
jcovariance, var (f
i) representation feature vector f
ivariance.According to calculating cross-correlation coefficient η, if η is <0.6, judge that current interpolation is characterized as effectively, otherwise judge that current signature is as invalid, need again choose arbitrary dimensional feature.Arbitrary dimensional feature f is added if continue again
i+1, need judge and current signature vector set { f
0, f
1... f
icross-correlation coefficient whether satisfy condition.
The generation of 2.2 cascaded thresholds, the soft cascade sorter h of the Linear SVM of structure n dimension
k(x):
Wherein w
ifor the support vector of categorised decision plane i, x
ifor the i dimensional feature of correspondence.According to arest neighbors feature, cascade classifier judges that whether it is effective, then according to h when every one-level Feature Selection
kx () calculates the response of all samples, find positive sample boundary to classify corresponding threshold value, and this grade of corresponding threshold value and feature will join k+1 level calculated response h
k+1(x).
The judgement of 2.3 cascade sorts.Window collection to be detected is sent into soft cascade sorter successively, the threshold value obtained according to upper step and feature are treated detection window and are carried out window screening, will non-targeted be thought when the response of current window is less than decision-making value, then remaining window be carried out next stage cascade sort judgement.If when current window response is higher than current level sorter decision-making value, then think that current window is target.
3. soft cascade sorter features training
3.1 positive samples generate, the window of the positive sample object image different scale demarcated is scanned, near target neighborhood, choose 10 encirclement frames nearest with it, the size intersecting area ratio according to window judges that whether the current positive sample chosen is eligible.It should be noted that, choose at most 4 location window under often kind of yardstick as positive sample, to ensure the completeness of training sample.
3.2 negative samples generate, while positive sample is chosen, translation is up and down carried out in positive sample window position, and when window intersects area lower than thinking during certain threshold value that the sample that current window produces is negative sample, often kind of yardstick produces maximum 4 location window equally as negative sample.
3.3 initial characteristics training, generate according to positive negative sample and obtain training sample set, each 200 of the positive negative sample of random selecting, normalize to 60 × 60 by unified for the lower sample of different scale, to sample travel direction gradient binary pattern feature interpretation, train preliminary classification device h
0(x).
3.4 cascade nature training, according to initial soft cascade, target detection is carried out to all images, according to positive sample window position, flase drop window is added next stage sorter features training, after completing image object detection, to newly obtaining positive and negative sample set re-training next stage classification h
k(x), and k ∈ [2 ,+∞), upgrade end condition until meet.
4. target window is followed the trail of and is upgraded
4.1 cascade target windows detect.According to threshold value and the character pair vector of the soft cascade SVM classifier of training, carry out the judgement of hierarchical classification device to image sequence, final cascade decision-making judges that output window is present frame target window.
4.2 extract target window unique point.Upper step obtains cascade classifier output window, utilizes shi-Tomasi angular-point detection method to extract the unique point of target window.It should be noted that the real-time in order to follow-up tracking and accuracy, the poorest quality assurance of Corner Detection is excellent, controls window interior angle spot check detecting number and is not more than 20.
4.3 target window tracking point screenings, Median-Flow tracker forward is utilized to track the (n+1)th frame according to the tracking point of current n-th frame window, backward tracing is to the n-th frame again, calculation window interior angle point number change and the Euclidean distance between back tracking point and former angle point, setting decision threshold screens best tracking point.
4.4 target window update strategies, utilize best trace point to calculate next frame target window predicted position, utilize the preliminary classification device of cascade classifier to carry out target discrimination.If current window is judged as target, think that current tracking is effective, otherwise need to carry out sliding window search according to cascade classifier, re-start target detection.
Below in conjunction with concrete scene, example test detailed is further done to the real-time target detection method based on direction gradient binary pattern and soft cascade SVM that the present invention proposes.At hardware platform Interi5+4GDDR3RAM, software platform OpenCV/C++ implements the example method, accompanying drawing 3 Scene multi-target detection is tested, and the present embodiment reaches 96% to target detection rate, and single frames working time is 27ms; In accompanying drawing 4, under complex scene, vehicle target detection is tested, and the present embodiment reaches 95.3% to target detection rate, single frames 25ms working time.
From above technical scheme, real-time target detection method based on direction gradient binary pattern and soft cascade SVM provided by the invention, the method is based on soft cascade Support Vector Machine SVM, adopt and describe in order to target signature based on direction gradient binary pattern feature, improve the robustness of feature interpretation to complex scene background and illumination variation; Utilize detected image random site to generate positive negative sample when carrying out features training, finally adopt shi-Tomasi Corner Detection extract minutiae to complete target tracking to upgrade, soft cascade reduces target window screening, and window follows the trail of the stability and real-time that improve multiframe Sequence Detection.Method strong robustness and the real-time of the present invention's proposition are high, and target detection performance is excellent.
According to the disclosure, also propose a kind of real-time target pick-up unit based on direction gradient binary pattern and soft cascade SVM, this device comprises:
For the first module of ORBP feature extraction, this first module is arranged for carries out pretreatment operation to image source sample, utilizes Sobel edge and local direction gradients setup ORBP feature;
For building second module of soft cascade sorter SVM, this second module is configured to utilize cross-correlation characteristic similarity to judge, and whether this sample characteristics is chosen effective, according to h
kx () calculates the response of all samples, find positive sample boundary to classify corresponding threshold value, and this grade of corresponding threshold value and feature will join k+1 level calculated response h
k+1(x); Then window collection to be detected is sent into soft cascade sorter successively, by judging that current window response judges whether to belong to target.；
For training the 3rd module of soft cascade sorter, the 3rd module is arranged for carrying out positive sample negative sample to the positive sample object image of demarcation and generates, and carries out ORBP feature interpretation to sample; Then initial sorter h is trained by SVM
0x (), carries out target detection checking according to initial sorter to sample image, be again updated to by negative sample in the training of next stage SVM cascade classifier, until complete the training of final cascade classifier;
The four module upgraded is followed the trail of for target window, this four module is configured to the sorter of training out according to soft cascade SVM, treat detected image sequence and carry out target window detection, utilize shi-Tomasi angular-point detection method to extract the unique point of target window, judge current signature point whether as best tracking point according to Median-Flow tracker; Then calculate next frame target window predicted position by best trace point, utilize the initial sorter of cascade classifier to carry out target discrimination, final output target detection window.
Be to be understood that, the first module that the present embodiment proposes, the second module, the 3rd module and four module, its function, effect and effect are illustrated in the description of the above real-time target detection method based on direction gradient binary pattern and soft cascade SVM, its implementation and done exemplary illustration in the aforementioned embodiment about real-time target detection method, does not repeat them here.
According to aforementioned embodiments of the present invention, such as based on the real-time target detection method of direction gradient binary pattern and soft cascade SVM and the real-time target pick-up unit based on direction gradient binary pattern and soft cascade SVM, the present invention also proposes a kind of for realizing the computer system detected based on the real-time target of direction gradient binary pattern and soft cascade SVM, and this computer system comprises:
Storer;
One or more processor;
One or more module, this one or more module is stored in which memory and is configured to be performed by described one or more processor, and described one or more module comprises following processing module:
For the first module of ORBP feature extraction, this first module is arranged for carries out pretreatment operation to image source sample, utilizes Sobel edge and local direction gradients setup ORBP feature;
For building second module of soft cascade sorter SVM, this second module is configured to utilize cross-correlation characteristic similarity to judge, and whether this sample characteristics is chosen effective, according to h
kx () calculates the response of all samples, find positive sample boundary to classify corresponding threshold value, and this grade of corresponding threshold value and feature will join k+1 level calculated response h
k+1(x); Then window collection to be detected is sent into soft cascade sorter successively, by judging that current window response judges whether to belong to target.；
For training the 3rd module of soft cascade sorter, the 3rd module is arranged for carrying out positive sample negative sample to the positive sample object image of demarcation and generates, and carries out ORBP feature interpretation to sample; Then initial sorter h is trained by SVM
0x (), carries out target detection checking according to initial sorter to sample image, be again updated to by negative sample in the training of next stage SVM cascade classifier, until complete the training of final cascade classifier;
The four module upgraded is followed the trail of for target window, this four module is configured to the sorter of training out according to soft cascade SVM, treat detected image sequence and carry out target window detection, utilize shi-Tomasi angular-point detection method to extract the unique point of target window, judge current signature point whether as best tracking point according to Median-Flow tracker; Then calculate next frame target window predicted position by best trace point, utilize the initial sorter of cascade classifier to carry out target discrimination, final output target detection window.
Although the present invention with preferred embodiment disclose as above, so itself and be not used to limit the present invention.Persond having ordinary knowledge in the technical field of the present invention, without departing from the spirit and scope of the present invention, when being used for a variety of modifications and variations.Therefore, protection scope of the present invention is when being as the criterion depending on those as defined in claim.
Claims (7)
1., based on a real-time target detection method of direction gradient binary pattern and soft cascade SVM, be characterised in that, comprise the following steps:
(1) ORBP feature extraction: carry out contrast variation's pre-service to image source sample, is divided into K part by gradient direction, respectively all directions block gradient figure under computed image Sobel edge; Then according to the cumulative response of direction block gradient figure level and vertical direction, ORBP feature is generated;
(2) structure of soft cascade sorter SVM: according to the ORBP feature of image source sample, calculate the response of all samples, the threshold value finding the classification of positive sample boundary corresponding and proper vector, then window to be detected is sent into soft cascade sorter successively, judge whether to belong to target by current window response magnitude;
(3) soft cascade sorter features training: positive negative sample generation is carried out to the positive sample object image of demarcation, the positive negative sample of random selecting is each N number of, ORBP feature interpretation is carried out to sample, then utilizes the soft cascade SVM classifier of structure to complete and sample characteristics is trained;
(4) target window is followed the trail of and is upgraded: the sorter of training out according to soft cascade SVM, target window detection is carried out to image sequence, utilize shi-Tomasi angular-point detection method to extract the unique point of target window, judge current signature point whether as best tracking point according to Median-Flow tracker; Then calculate next frame target window predicted position by best trace point, utilize the initial sorter of cascade classifier to carry out target discrimination, final output target detection window.
2., as claimed in claim 1 based on the real-time target detection method of direction gradient binary pattern and soft cascade SVM, it is characterized in that, in described step (1), the concrete grammar of ORBP feature extraction comprises the following steps:
(1) carry out contrast variation's pretreatment operation to image source sample, pretreatment operation comprises Gaussian smothing filtering and contrast normalized;
(2) gradient direction is divided into K part, the gradient magnitude respectively under calculating Sobel edge under each gradient direction, puts into corresponding M sub-block matrix by long for the gradient-norm of K direction scope, generate corresponding edge direction block gradient figure;
(3) carry out level to each edge direction gradient map and vertical direction edge divides, level of adding up respectively adds up to respond with vertical direction;
(4) according to horizontal direction up and down accumulative response with add up response magnitude about vertical direction and compare and generate ORBP feature.
3. as claimed in claim 1 based on the real-time target detection method of direction gradient binary pattern and soft cascade SVM, it is characterized in that, the structure of soft cascade SVM in described step (2), specific implementation comprises:
(1) calculate cross-correlation characteristic similarity to judge: to image source sample extraction ORBP feature, utilize the similarity between cross-correlation calculation feature, judge whether effective this sample characteristics is chosen according to normalized-cross-correlation function;
(2) generation of cascaded thresholds: the soft cascade sorter h of the Linear SVM of structure n dimension
k(x):
Wherein w
ifor the support vector of categorised decision plane i, x
ifor the i dimensional feature of correspondence; According to cross-correlation feature, cascade classifier judges that whether it is effective, then according to h when every one-level Feature Selection
kx () calculates the response of all samples, find positive sample boundary to classify corresponding threshold value, and this grade of corresponding threshold value and feature will join k+1 level calculated response h
k+1(x);
(3) judgement of cascade sort: all windows to be detected are sent into soft cascade sorter successively, the threshold value utilizing cascade to obtain and feature are treated detection window and are carried out window screening, non-targeted will be thought when the response of current window is less than decision-making value, then remaining window is carried out the judgement of next stage cascade classifier: if when current window response is higher than current level sorter decision-making value, then think that current window is target.
4., as claimed in claim 1 based on the real-time target detection method of direction gradient binary pattern and soft cascade SVM, it is characterized in that, in described step (3), generate positive negative sample and specifically comprise:
1) positive sample generates: scan the window of the positive sample object image different scale demarcated, near target neighborhood, choose 10 encirclement frames nearest with it, the size intersecting area ratio according to window judges that whether the current positive sample chosen is eligible;
2) negative sample generates: while positive sample is chosen, translation is up and down carried out in positive sample window position, when window intersects area lower than thinking during certain threshold value that the sample that current window produces is negative sample, often kind of yardstick produces maximum 4 location window equally as negative sample.
5., as claimed in claim 3 based on the real-time target detection method of direction gradient binary pattern and soft cascade SVM, it is characterized in that, described cross-correlation characteristic similarity judge concrete grammar as:
ORBP feature extraction is carried out to image source sample, optionally gets one-dimensional characteristic as initial characteristics f
0, add other one-dimensional characteristics as second feature f
1, calculate feature f
0with f
1normalized-cross-correlation function η, the computing method of normalized-cross-correlation function η as shown in the formula:
Wherein, f
irepresent the i-th dimensional feature vector, cov (f
i, f
j) representation feature vector f
iwith f
jcovariance, var (f
i) representation feature vector f
ivariance; According to calculating cross-correlation coefficient η, if η < is η
0then judge that current interpolation is characterized as effectively, otherwise judge that current signature is as invalid, need again choose arbitrary dimensional feature; Arbitrary dimensional feature f is added if continue again
i+1, need judge and current signature vector set { f
0, f
1... f
icross-correlation coefficient whether satisfy condition.
6., based on a real-time target pick-up unit of direction gradient binary pattern and soft cascade SVM, it is characterized in that, this device comprises:
For the first module of ORBP feature extraction, this first module is arranged for carries out pretreatment operation to image source sample, utilizes Sobel edge and local direction gradients setup ORBP feature;
For building second module of soft cascade sorter SVM, this second module is configured to utilize cross-correlation characteristic similarity to judge, and whether this sample characteristics is chosen effective, according to h
kx () calculates the response of all samples, find positive sample boundary to classify corresponding threshold value, and this grade of corresponding threshold value and feature will join k+1 level calculated response h
k+1(x); Then window collection to be detected is sent into soft cascade sorter successively, by judging that current window response judges whether to belong to target;
For training the 3rd module of soft cascade sorter, the 3rd module is arranged for carrying out positive sample negative sample to the positive sample object image of demarcation and generates, and carries out ORBP feature interpretation to sample; Then initial sorter h is trained by SVM
0x (), carries out target detection checking according to initial sorter to sample image, be again updated to by negative sample in the training of next stage SVM cascade classifier, until complete the training of final cascade classifier;
The four module upgraded is followed the trail of for target window, this four module is configured to the sorter of training out according to soft cascade SVM, treat detected image sequence and carry out target window detection, utilize shi-Tomasi angular-point detection method to extract the unique point of target window, judge current signature point whether as best tracking point according to Median-Flow tracker; Then calculate next frame target window predicted position by best trace point, utilize the initial sorter of cascade classifier to carry out target discrimination, final output target detection window.
7., for realizing the computer system detected based on the real-time target of direction gradient binary pattern and soft cascade SVM, it is characterized in that, this computer system comprises:
Storer;
One or more processor;
One or more module, this one or more module is stored in which memory and is configured to be performed by described one or more processor, and described one or more module comprises following processing module:
For the first module of ORBP feature extraction, this first module is arranged for carries out pretreatment operation to image source sample, utilizes Sobel edge and local direction gradients setup ORBP feature;
For building second module of soft cascade sorter SVM, this second module is configured to utilize cross-correlation characteristic similarity to judge, and whether this sample characteristics is chosen effective, according to soft cascade sorter h
kx () calculates the response of all samples, find positive sample boundary to classify corresponding threshold value, and this grade of corresponding threshold value and feature will join k+1 level calculated response h
k+1(x); Then window collection to be detected is sent into soft cascade sorter successively, by judging that current window response judges whether to belong to target;
For training the 3rd module of soft cascade sorter, the 3rd module is arranged for carrying out positive sample negative sample to the positive sample object image of demarcation and generates, and carries out ORBP feature interpretation to sample; Then initial sorter h is trained by SVM
0x (), carries out target detection checking according to initial sorter to sample image, be again updated to by negative sample in the training of next stage SVM cascade classifier, until complete the training of final cascade classifier;
The four module upgraded is followed the trail of for target window, this four module is configured to the sorter of training out according to soft cascade SVM, treat detected image sequence and carry out target window detection, utilize shi-Tomasi angular-point detection method to extract the unique point of target window, judge current signature point whether as best tracking point according to Median-Flow tracker; Then calculate next frame target window predicted position by best trace point, utilize the initial sorter of cascade classifier to carry out target discrimination, final output target detection window.
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 true CN105260749A (en) | 2016-01-20 |
| CN105260749B 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) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106228162A (en) * | 2016-07-22 | 2016-12-14 | 王威 | A kind of quick object identification method of mobile robot based on degree of depth study |
| CN106251362A (en) * | 2016-07-15 | 2016-12-21 | 中国电子科技集团公司第二十八研究所 | A kind of sliding window method for tracking target based on fast correlation neighborhood characteristics point and system |
| CN107133558A (en) * | 2017-03-13 | 2017-09-05 | 北京航空航天大学 | A kind of infrared pedestrian's conspicuousness detection method based on probability propagation |
| CN108073940A (en) * | 2016-11-18 | 2018-05-25 | 北京航空航天大学 | A kind of method of 3D object instance object detections in unstructured moving grids |
| 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 |
| CN108960320A (en) * | 2018-06-26 | 2018-12-07 | 湖北工业大学 | Train coupling cock fault picture real-time detection method |
| CN109544594A (en) * | 2017-09-22 | 2019-03-29 | 财团法人资讯工业策进会 | Target tracking method and system under multiple nonlinear distorted lenses |
| CN110335252A (en) * | 2019-06-04 | 2019-10-15 | 大连理工大学 | Picture quality detection method based on background characteristics point motion analysis |
| CN110852323A (en) * | 2019-11-18 | 2020-02-28 | 南京莱斯电子设备有限公司 | Angular point-based aerial target detection method |
| CN111652292A (en) * | 2020-05-20 | 2020-09-11 | 贵州电网有限责任公司 | Similar object real-time detection method and system based on NCS and MS |
| CN111681262A (en) * | 2020-05-08 | 2020-09-18 | 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 |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070112699A1 (en) * | 2005-06-20 | 2007-05-17 | Samsung Electronics Co., Ltd. | Image verification method, medium, and apparatus using a kernel based discriminant analysis with a local binary pattern (LBP) |
| US20080166026A1 (en) * | 2007-01-10 | 2008-07-10 | Samsung Electronics Co., Ltd. | Method and apparatus for generating face descriptor using extended local binary patterns, and method and apparatus for face recognition using extended local binary patterns |
| 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 |
-
2015
- 2015-11-02 CN CN201510733481.1A patent/CN105260749B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070112699A1 (en) * | 2005-06-20 | 2007-05-17 | Samsung Electronics Co., Ltd. | Image verification method, medium, and apparatus using a kernel based discriminant analysis with a local binary pattern (LBP) |
| US20080166026A1 (en) * | 2007-01-10 | 2008-07-10 | Samsung Electronics Co., Ltd. | Method and apparatus for generating face descriptor using extended local binary patterns, and method and apparatus for face recognition using extended local binary patterns |
| 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 |
|---|
| 李星等: "基于HOG特征和SVM的前向车辆识别方法", 《计算机科学》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106251362A (en) * | 2016-07-15 | 2016-12-21 | 中国电子科技集团公司第二十八研究所 | A kind of sliding window method for tracking target based on fast correlation neighborhood characteristics point and system |
| 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 |
| CN106228162A (en) * | 2016-07-22 | 2016-12-14 | 王威 | A kind of quick object identification method of mobile robot based on degree of depth study |
| CN108073940A (en) * | 2016-11-18 | 2018-05-25 | 北京航空航天大学 | A kind of method of 3D object instance object detections in unstructured moving grids |
| 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 |
| CN107133558A (en) * | 2017-03-13 | 2017-09-05 | 北京航空航天大学 | A kind of infrared pedestrian's conspicuousness detection method based on probability propagation |
| CN109544594B (en) * | 2017-09-22 | 2023-05-09 | 财团法人资讯工业策进会 | Target tracking method and system under multiple nonlinear distortion lenses |
| CN109544594A (en) * | 2017-09-22 | 2019-03-29 | 财团法人资讯工业策进会 | Target tracking method and system under multiple nonlinear distorted 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 |
| CN108960320A (en) * | 2018-06-26 | 2018-12-07 | 湖北工业大学 | Train coupling cock fault picture real-time detection method |
| 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 |
| CN110335252A (en) * | 2019-06-04 | 2019-10-15 | 大连理工大学 | Picture quality detection method based on background characteristics point motion analysis |
| CN110852323A (en) * | 2019-11-18 | 2020-02-28 | 南京莱斯电子设备有限公司 | Angular point-based aerial target detection method |
| CN111681262A (en) * | 2020-05-08 | 2020-09-18 | 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 |
| 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 |
| CN111652292A (en) * | 2020-05-20 | 2020-09-11 | 贵州电网有限责任公司 | Similar object real-time detection method and system based on NCS and MS |
| 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 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105260749B (en) | 2018-11-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105260749A (en) | Real-time target detection method based on oriented gradient two-value mode and soft cascade SVM | |
| CN109919974B (en) | Online multi-target tracking method based on R-FCN frame multi-candidate association | |
| CN102306290B (en) | Face tracking recognition technique based on video | |
| US10198657B2 (en) | All-weather thermal-image pedestrian detection method | |
| CN103246896B (en) | A kind of real-time detection and tracking method of robustness vehicle | |
| CN102998316B (en) | Transparent liquid impurity detection system and detection method thereof | |
| CN103824070A (en) | Rapid pedestrian detection method based on computer vision | |
| CN106022231A (en) | Multi-feature-fusion-based technical method for rapid detection of pedestrian | |
| CN102147869A (en) | Pedestrian detection method based on foreground analysis and pattern recognition | |
| CN106778570B (en) | A kind of pedestrian detection and tracking in real time | |
| CN104091171A (en) | Vehicle-mounted far infrared pedestrian detection system and method based on local features | |
| CN104615986A (en) | Method for utilizing multiple detectors to conduct pedestrian detection on video images of scene change | |
| CN103593679A (en) | Visual human-hand tracking method based on online machine learning | |
| CN103955949A (en) | Moving target detection method based on Mean-shift algorithm | |
| Kim et al. | Autonomous vehicle detection system using visible and infrared camera | |
| CN109101932A (en) | The deep learning algorithm of multitask and proximity information fusion based on target detection | |
| CN107092878A (en) | It is a kind of based on hybrid classifer can autonomous learning multi-target detection method | |
| Su et al. | A new local-main-gradient-orientation HOG and contour differences based algorithm for object classification | |
| CN101470802A (en) | Object detection apparatus and method thereof | |
| CN104123714A (en) | Optimal target detection scale generation method in people flow statistics | |
| Yin | Object Detection Based on Deep Learning: A Brief Review | |
| CN104318589A (en) | ViSAR-based anomalous change detection and tracking method | |
| CN112347967B (en) | Pedestrian detection method fusing motion information in complex scene | |
| CN103996207A (en) | Object tracking method | |
| CN111275733A (en) | Method for realizing rapid tracking processing of multiple ships based on deep learning target detection technology |
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 |