CN101930549B - Second generation curvelet transform-based static human detection method - Google Patents
Second generation curvelet transform-based static human detection method Download PDFInfo
- Publication number
- CN101930549B CN101930549B CN 201010259562 CN201010259562A CN101930549B CN 101930549 B CN101930549 B CN 101930549B CN 201010259562 CN201010259562 CN 201010259562 CN 201010259562 A CN201010259562 A CN 201010259562A CN 101930549 B CN101930549 B CN 101930549B
- Authority
- CN
- China
- Prior art keywords
- sample
- human body
- training
- feature vector
- image
- 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.)
- Expired - Fee Related
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a second generation curvelet transform-based static human detection method, which mainly solves the problem of high detection false-alarm rate in the conventional human detection technology. The detection process comprises the following steps of: acquiring a negative sample through bootstrap operation of the negative sample, and forming a training sample set by the negative sample and other positive samples in a database; calculating curvelet transform-based feature vectors of all training samples to form a training sample feature set; performing classification training on the sample feature set by adopting an AdaBoost algorithm to obtain a classifier; inputting a to-be-tested image with any size, calculating curvelet transform-based feature vectors of all scan window images in the to-be-tested image; inputting the curvelet transform-based feature vectors of all scan window images into the obtained classifier for classification; and according to classification results, combining all scan windows classified as human by utilizing a main window merging method to form the final human detection result. The method has the advantages of high detection accuracy and low false-alarm rate, and can be used for classifying and detecting human in the image.
Description
Technical field
The invention belongs to mode identification technology, relate to human body detecting method, can be used for the classification and detection to the human body in the image and other complex target.
Background technology
Human detection has many important application in computer vision, such as video monitoring, intelligent automobile and intelligent transportation, robot and senior man-machine interaction etc.Yet because the impact of the factors such as the diversity of the variation of human body self attitude, clothes and illumination, the cosmetic variation of human body is very large, causes human detection to become a very hard problem.
At present, in the still image method of human detection mainly contain method based on manikin, based on the method for template matches with based on the method for statistical classification.Method based on manikin has clear and definite model, can process occlusion issue, and can infer the attitude of human body.Shortcoming is that the structure of model is relatively more difficult, finds the solution also more complicated.Calculate simply based on the method for template matches, shortcoming is because the complicacy of human body attitude is difficult to construct enough templates to process different attitudes.Method based on statistical classification obtains a sorter by machine learning from a series of training data learnings, represents human body with this sorter, then utilizes this sorter input window is classified and to identify.Advantage based on the method for statistical classification is the comparison robust, and shortcoming is to need a lot of training datas, and is difficult to the problem that solves attitude and block.
Method based on statistical classification mainly comprises two steps: feature extraction and classifier design.Wherein the sign ability of selected feature will directly affect the performance of sorter.Present selected feature comprises: original gray feature space, Haar wavelet character, shape description symbols feature, Gabor feature, oriented histogram of gradients HOG feature and SIFT feature etc.Wherein the extraction rate of HOG feature extracting method is fast, and it is also higher to detect accuracy, thus, is widely used in recent years.How further to improve the accuracy that detects, become one of key problem of this research field.
According to the method for designing of sorter, existing method based on statistical classification can be divided into method based on neural network NN, based on the method for support vector machines with based on the method for AdaBoost.AdaBoost is a kind of Boosting algorithm with adaptivity, it is by setting up many learning machine combinations so that the performance of weak learning machine gets a promotion, because what it was exclusive pays close attention to the self-adaptation of learning machine performance with to excessively learning the immunity of phenomenon, having caused in recent years widely.
The method that classical characteristics of human body extracts is the method for the histograms of oriented gradients HOG of Dalal and Trigg proposition, and their result shows that HOG method method before is greatly improved in the detection accuracy; Another method is the edgelet method, and it utilizes the edge direction characteristic of a kind of edgelet of being called as descriptor; Edgelet is exactly one section straight line or the curve on the detected edge; Granularity is adjustable, and gradient subregion GGP descriptor has merged heterogeneous feature, and the representation of concept by granularity out.Because when having Clutter edge in background, it is poor that the HOG feature shows, and has the situation that the feature in some backgrounds is judged as the characteristics of human body, the false alarm rate that therefore detects is higher.
Candes and Donoho have introduced a kind of new multiscale analysis system and have been called the curve wave conversion.The curve wave conversion is a multiple dimensioned pyramid, at each yardstick a lot of directions and position is arranged.The curve wave energy is the uncontinuity at edge in the presentation video effectively.
The people such as MohamedElAroussi have been applied to the method for curve wave conversion in the detection of people's face, the method for detecting human face of block-based curve wave conversion has been proposed, the statistical nature that utilization is extracted from the piece that the coefficient of curve wave conversion is divided carries out people's face as proper vector and detects, at ORL, obtain good detection effect on YALE and the FERET data set, but gone back at present the detection that nobody is used for the curve wave conversion human body.
Summary of the invention
The object of the invention is to overcome the high problem of above-mentioned HOG method false alarm rate when having noisy background, the human body detecting method that a kind of edge based on the curve wave conversion and texture union feature extract has been proposed, with the false alarm rate of reduction human detection, thereby improved the accuracy that detects.
The know-why that realizes the object of the invention is that the method that the people's face with above-mentioned block-based curve wave conversion detects is applied in the human detection, and above-mentioned method improved, extracting method based on the human body edge feature of curve wave conversion has been proposed, and added the textural characteristics that extracts the coefficient behind the curve wave conversion, extract the feature of human body training sample set, by these features of AdaBoost Algorithm for Training, obtain sorter, and use the method realization of sliding window scanning to the detection of human body in the still image.Detailed process is as follows:
(1) in the INRIA database, obtains negative sample by bootstrapping operation, these negative samples other positive sample composing training sample set in database;
(2) extract the edge feature vector based on the curve wave conversion that training sample is concentrated each training sample, and the edge feature vector that extracts has been carried out the selection of optimum statistic;
(3) extract the texture feature vector based on the curve wave conversion that training sample is concentrated each training sample, and with the vectorial cascade of edge feature that this texture feature vector and step (2) are extracted, jointly consist of characteristics of human body's vector;
(4) characteristics of human body of all training samples vector in the calculation training sample set forms the sample characteristics collection, utilizes the AdaBoost sorting algorithm that it is carried out classification based training, obtains a sorter;
(5) input arbitrary size by altimetric image, adopt the method for sliding window scanning to being scanned detection by altimetric image, and the characteristics of human body based on the curve wave conversion who calculates all scanning window images is vectorial, is input in the step (4) and classifies in the resulting sorter;
(6) according to the classification results of sorter output, utilize the main window act of union, all scanning windows that are divided into human body are made up, form final human detection result.
The present invention has the following advantages compared with prior art:
1, because the present invention has used two kinds of different feature extracting methods to carry out the bootstrapping operation of negative sample, so that the negative sample that obtains is more representative, the sorter of the sample set of its formation training has better classification performance.
2, because the statistical nature of coefficient behind the curve wave conversion that the present invention uses when the edge feature that extracts based on the curve wave conversion, can extract compactness and the edge feature of the meaning of representing is arranged, reduced simultaneously the dimension of feature, can be more conducive to the sorter training, the further feature close with intrinsic dimensionality compared, and corresponding sorter reaches the required training time of identical training error to be shortened greatly.
3, since the statistical information that the co-occurrence matrix of utilization of the present invention coefficient behind the curve wave conversion extracts as textural characteristics, expanded edge feature, thereby further improved the accuracy that detects, the false alarm rate high shortcoming that has detected when having remedied HOG noisy for background.
4, experimental result shows, use among the present invention based on the better picture engraving of the edge of curve wave conversion and texture union feature so that compare with traditional HOG feature extracting method, be significantly improved at classification accuracy rate, it is also more accurate to detect.
Description of drawings
Fig. 1 is schematic flow sheet of the present invention;
Fig. 2 is the positive sample of part and the negative sample image that uses among the present invention;
Fig. 3 is the present invention and traditional HOG feature extracting method and based on the Edge Gradient Feature method of the curve wave conversion error attenuated comparison diagram when the sample training;
Fig. 4 is the present invention and traditional HOG feature extracting method and based on the recipient operating characteristic curve ROC of Edge Gradient Feature method when test sample book is classified of curve wave conversion;
Fig. 5 is that some when test sample book is classified of the sorter of method of the present invention training are by the test sample book of mis-classification;
Fig. 6 is the human detection result figure that the present invention is used for still image.
Embodiment
The present invention utilizes the curve wave conversion, extracts the edge feature associating textural characteristics of curve wave conversion coefficient as Characteristic of Image, classify and detected image in human body.Behind the edge and texture union feature that extract based on the curve wave conversion, utilize the AdaBoost sorting algorithm to carry out sample training, and classification results and HOG feature are compared.Describe in detail such as Fig. 1, Fig. 3 and Fig. 4.
With reference to Fig. 1, specific implementation process of the present invention is as follows:
The database that the present invention uses is from INRIA somatic data storehouse, and download address is: http://pascal.inrialpes.fr/data/human/.Because this database do not provide enough negative samples, thus need to be take this database as the basis, and the bootstrapping by negative sample operates to obtain negative sample.
The present invention is in order to make training sample have more representativeness, used respectively the HOG feature and based on edge and these two kinds of features of texture union feature of curve wave conversion, carry out the bootstrapping operation of negative sample when feature extraction, can obtain two negative sample collection; Then getting it occurs simultaneously as final negative sample collection; Added by this negative sample collection at last positive sample set in the INRIA database to obtain final training sample set, have 6246 samples, as shown in Figure 2, wherein Fig. 2 (a) be just sample of part, and Fig. 2 (b) is the part negative sample; As training set, 1132 positive samples and 821 negative samples are as test set with wherein 2416 positive samples and 1877 negative samples, and size is 64 * 128 pixels.
The bootstrapping operating process of described negative sample is as follows:
(1.1) from the INRIA database, appoint first and get the positive sample of sub-fraction and negative sample, carry out feature extraction, training preliminary classification device;
(1.2) use this preliminary classification device, the non-human body image of in the Test database all the other, to these non-human body images, be divided into a random choose part in the scanning window image of positive sample in mistake and form new negative sample collection with current negative sample, use random choose can avoid sneaking into the sample image of a large amount of feature similarities;
(1.3) repeat (the 1.1)-feature extraction of (1.2), training classifier, the non-human body image of detection and form new this process of negative sample collection until collect final negative sample.
Step 2 is utilized the curve wave conversion, calculates the edge feature vector based on the curve wave conversion of all training samples, and the edge feature vector that extracts is carried out the selection of optimum statistic;
(2.1) each training sample is done fast discrete curve wave conversion, the parameter of curve wave conversion is set to real-valued curve wave conversion; And scale parameter is 3 layers; Inferior rough layer has 8 directional subbands; Detailed level is curve wave system number, has 16 directional subbands;
(2.2) for the coefficient after the conversion, at first the sub-band coefficients matrix that has formed objects in inferior rough layer and the detailed level is spliced, then the matrix of coefficients of every one deck is carried out two according to 8 * 8 coefficient magnitude and advance to deduct marks, obtain curve ripple piece, do not have overlapping in the subdivision process between the piece;
(2.3) in each piece, extract these several statistics of energy, entropy, standard deviation, average, maximal value, minimum value and contrast, and the resulting statistic of different piecemeals is connect the initial edge proper vector that obtains image; And the initial edge proper vector has been carried out the selection of optimum statistic;
(2.4) calculate the initial edge proper vector of all training samples, obtain the initial edge proper vector of training sample set, put into and carry out training classifier in the AdaBoost algorithm, because the selected probability of feature is directly proportional with the sign ability of feature, so during statistics training AdaBoost sorter, with every kind of selected number of times of statistic the statistic in the step (2.3) is screened, just can obtain optimum statistic.The experimental result that optimum statistic is selected is as shown in table 1:
Table 1 optimum statistic choice experiment result
Wherein, C is sorter exercise wheel number,
The present invention is according to the statistic choice experiment, preferentially chosen following several statistic as optimum statistic:
Energy:
Entropy:
Contrast:
Standard deviation:
Maximal value: max (c
I, j), wherein, c
I, jBe the coefficient take (i, j) as coordinate in each curve ripple piece, N is the number of contained element in each curve ripple piece;
(2.5) optimum statistic with the different curve ripple pieces of training sample connects, and forms a proper vector, i.e. the edge feature vector.
Step 3 is extracted the texture feature vector based on the curve wave conversion that training sample is concentrated each training sample, and with the vectorial cascade of edge feature that this texture feature vector and step 2 are extracted, jointly consists of characteristics of human body's vector.
(3.1) take 1/2 overlapping mode to carry out subdivision with the piece of 8*8 for the sub-band coefficients of the rough layer after the conversion;
(3.2) its co-occurrence matrix is calculated in the coefficient fritter of each 8 * 8 size, the quantification progression of this co-occurrence matrix is 16;
(3.3) for the co-occurrence matrix of the coefficient fritter of each 8 * 8 size, calculate its angle second moment, entropy, contrast, correlativity, average and and variance and this six statistics, that is:
The angle second moment:
Entropy:
Contrast:
N=|i-j|. wherein
Correlativity:
μ wherein
x, μ
y, σ
x, σ
yBe defined as respectively:
Average and:
Wherein:
Variance and:
Wherein, i represents the line number of element in the coefficient co-occurrence matrix of piece, and j represents the columns of element in the coefficient co-occurrence matrix of piece,
(i, j) individual element of coefficient co-occurrence matrix of expression piece, G represent the quantification progression of co-occurrence matrix of the rough layer sub-band coefficients subdivision sub-block of curve wave conversion;
(3.4) statistic with the co-occurrence matrix of the coefficient of different masses connects the texture feature vector that obtains image;
(3.5) texture feature vector of image and the edge feature vector of step 2 gained are carried out cascade as the proper vector of human body.
Step 4, the characteristics of human body of all training samples vector forms the training sample feature set in the calculation training sample set, utilizes the AdaBoost sorting algorithm that it is carried out classification based training, obtains a sorter;
(4.1) extract the edge feature vector sum texture feature vector that training sample is concentrated each sample according to the method for step 2 and step 3 respectively, and the edge feature vector sum texture feature vector of each training sample that will extract carries out cascade, obtains characteristics of human body's vector of each training sample;
(4.2) suppose that characteristics of human body's vector is the vector of a M dimension, it is N that training sample is concentrated the number of training sample, the proper vector of then training sample set being extracted will form the matrix of a N * M dimension, as the input of AdaBoost algorithm, train a sorter.
Described AdaBoost is a kind of Boosting algorithm with adaptivity, and it has adaptive learning machine performance and immune to crossing the study phenomenon by setting up many learning machine combinations so that the performance of weak learning machine gets a promotion.
In the Adaboost algorithm, represent the size that this sample is divided by mistake with the weighted value of each sample.In the weight renewal process that each is taken turns, can be become greatly by the weight of wrong minute sample, if a sample has been divided many times by mistake, the weight of this sample is just increasing so.
Step 5, the input arbitrary size by altimetric image, the method that adopts sliding window scanning is to being scanned detection by altimetric image, and calculates the characteristics of human body's vector based on the curve wave conversion of all scanning window images, the specific implementation step is as follows:
(5.1) the input arbitrary size by altimetric image, be that the ratio of [0.4,0.5,0.6,0.7,0.8,0.9] is carried out convergent-divergent with it by pantograph ratio;
(5.2) will in behind the described ratio scaling by the zone of a sample size in the altimetric image upper left corner as first scanning window, every to 8 pixels of right translation or downwards 8 pixels of translation obtain thus one group of scanning window as a new scanning window;
(5.3) image section in the scanning window is carried out the curve wave conversion, matrix of coefficients for the curve wave conversion, after the splicing of matrix of coefficients piece, to the matrix of coefficients on each yardstick carry out respectively 8 * 8 two advance to deduct marks, calculate these a few class statistics of energy, entropy, contrast, standard deviation and maximal value on each curve ripple fritter, and the statistic of different masses is connect the edge feature vector that obtains last scanning window image;
(5.4) the overlapping mode of the coefficients by using 1/2 of the rough layer of matrix of coefficients of curve wave conversion is carried out subdivision with the fritter of 8*8; Calculate each fritter angle second moment, entropy, contrast, correlativity, average and and variance and this several statistics, and the statistic of different fritters connect the texture feature vector that obtains the scanning window image;
(5.5) the edge feature vector sum texture feature vector of scanning window image is carried out cascade and obtain final scanning window Characteristic of Image vector;
(5.6) proper vector of all scanning windows is input in the resulting sorter of step 4 classifies, obtain one group of result of determination.
Sorter will be judged the image human body whether in all scanning windows.Each scanning window can obtain one and judge mark, be the human body window if should judge mark greater than 0 this scanning window of expression, otherwise be non-human body window, with the sorter mark of the position of tested imagezoom ratio, all scanning window images and gained thereof the classification results as sorter output.
Step 6, the classification results according to sorter output utilizes the main window act of union, and all scanning windows that are divided into human body are made up, and forms final human detection result, and concrete steps are as follows:
(6.1) according to the fractional value of the scanning window image of sorter output, whether judgement is comprised human body in the altimetric image, if the scanning window of sorter output does not have the human body window, then do not comprised human body in the altimetric image, otherwise, from all human body windows, find out its sorter mark the highest as main window;
(6.2) main window and other human body windows are made up judgement, when other human body windows are in around the main window and overlapping being judged to less than 1/2 time do not made up, otherwise make up;
(6.3) with the border average of main window and the human body window that all need to make up as a testing result, the present invention, when asking the border average, with the sorter mark of the human body window weighted value as the border: the sorter mark is larger, its border weighting is larger, and is larger on final testing result impact; The sorter mark is less, and its border weighting is less, and is also less on final testing result impact, can make like this detection position of human body more accurate;
(6.4) the human body window of deletion main window and all participation combinations;
(6.5) if also have remaining human body window, then find out again sorter mark wherein the highest as main window, and repeat the operation of (2)-(4);
(6.6) mark all testing results at tested person's volume image, by the final human detection result of altimetric image, generally adopt rectangle frame to represent testing result as this, the human body that is detected is in the rectangle frame.
Effect of the present invention can further specify by following emulation:
1. emulation content:
Adopt feature extracting method and now widely used HOG feature extracting method based on the curve wave conversion proposed by the invention to carry out simulation comparison experiment.Positive sample set is wherein taken from the INRIA database, the negative sample collection is take this database as the basis, obtain by negative sample bootstrapping operation, the sample set that obtains at last has 6246 samples, wherein 2416 positive samples and 1877 negative samples are as training set, 1132 positive samples and 821 negative samples are as test set, and sample size is 64 * 128 pixels.Fig. 2 has provided part sample image wherein, and wherein Fig. 2 (a) is the positive sample of part, and Fig. 2 (b) is the part negative sample.
For every kind of feature extracting method, extract first the feature of training set, re-use the AdaBoost algorithm and carry out classification based training, obtain a sorter, then use this sorter that test set is carried out class test, at last input detects whether to comprise human body arbitrarily by altimetric image, if comprise human body then will detect the position of human body.Be 480 * 640 pixels or 640*480 pixel size by the altimetric image size wherein, scanning window is 64 * 128 pixels, and zoom factor is got [0.4,0.5,0.6,0.7,0.8,0.9].During scan image, use dense scanning, 8 pixels of directions X translation, 8 pixels of Y-direction translation.At first for the convergent-divergent yardstick, the scanning entire image, then therefrom extract the proper vector on all scanning window images, carry out Classification and Identification, the group of windows that will be divided at last human body is synthesized testing result, and is showed by altimetric image former, and what wherein window combination adopted is the method that main window merges, hardware platform is: Intel Core2 Duo CPU E6550@2.33GHZ, 2GB RAM, software platform is MATLAB 7.2.
2. simulation result and analysis:
Fig. 3 has provided the present invention and the error attenuated comparison diagram of traditional HOG feature extracting method when sample training.Wherein C is the exercise wheel number, and Error is training error.As can be seen from Figure 3, adopt based on the edge of curve wave conversion and the texture union feature error when the training will be all the time error when using the HOG feature and only adopting edge feature training based on the curve wave conversion.As seen, use edge and texture union feature based on the curve wave conversion to be more conducive to carry out the sorter training, corresponding sorter reaches the required training time of identical training error to be shortened greatly.
Table 2 has provided the present invention and traditional HOG and has only adopted the performance comparison result of Edge Gradient Feature method when test sample book is classified based on the curve wave conversion.
Fig. 4 has provided the recipient's operating characteristic curve ROC that with the sorter of three kinds of method training test sample book is classified respectively.As can be seen from Figure 4 the edge feature method HOG method recall rate in the situation of identical false alarm rate before based on the curve wave conversion improves, and method of the present invention, namely add after the texture information, recall rate improves again, has embodied its advantage on human detection.
Fig. 5 has provided some by the image of false retrieval, wherein Fig. 5 (a) is become the positive sample of negative sample by false retrieval, be owing to block, the variation of illumination and the reasons such as variation of attitude cause, Fig. 5 (b) is become the negative sample of positive sample by false retrieval, be the situation of false-alarm, because profile or the textural characteristics of some images are a bit similar to human body.
The contrast of table 2. classification results
100 training errors of taking turns just can be found out from table 2, use feature extracting method of the present invention, are more conducive to carry out the sorter training.Simultaneously, contrasting under the identical training error different sorters can find out the classification results of test set, feature extracting method of the present invention is compared the HOG feature extracting method, is enhanced at classification accuracy rate, and false alarm rate also has certain reduction with respect to the HOG method.In the table 2, last classifies the averaging time that each sample characteristics extracts as, and every width of cloth Characteristic of Image is slightly more longer than HOG method on the used time extracting, owing to carry out the curve wave conversion to image, and extract some statistics.
Fig. 6 has provided method of the present invention for the testing result of some human body image.Wherein Fig. 6 (a) is the result to the identification of scanning window Images Classification, the final human detection result of Fig. 6 (b) for the scanning window that is judged as human body in the image 6 (a) is made up, Fig. 6 (c) is the testing result of another width of cloth image.As can be seen from Figure 6, use the edge based on the curve wave conversion of the present invention and texture union feature extracting method, can from tested person's volume image, detect human body accurately, especially be greatly improved at the reduction false alarm rate.
Experiment shows, uses HOG feature extracting method and edge and texture union feature extracting method based on the curve wave conversion of the present invention, all can detect human body more accurately from tested person's volume image.But feature extracting method of the present invention has higher classification accuracy rate, has overcome the high shortcoming of false alarm rate of the existing detection of HOG feature extracting method, has improved the accuracy that detects, and is very suitable for the human detection of still image.
Claims (6)
1. static human detection method based on second generation curve wave conversion comprises following process:
(1) in the INRIA database, obtains negative sample by bootstrapping operation, these negative samples other positive sample composing training sample set in database;
(2) extract the edge feature vector based on second generation curve wave conversion that training sample is concentrated each training sample, and the edge feature vector that extracts has been carried out the selection of optimum statistic;
(3) extract the texture feature vector based on second generation curve wave conversion that training sample is concentrated each training sample, and with the vectorial cascade of edge feature that this texture feature vector and step (2) are extracted, jointly consist of characteristics of human body's vector;
(4) training sample is concentrated the edge feature vector sum texture feature vector of all training samples form the sample characteristics collection, utilized sample characteristics set pair AdaBoost sorter to carry out classification based training, obtain a sorter;
(5) the input arbitrary size by altimetric image, the method that adopts sliding window scanning is to being scanned detection by altimetric image, and calculates characteristics of human body's vector of all scanning window images, is input in the step (4) and classifies in the resulting sorter;
(6) according to the classification results of sorter output, utilize the main window act of union, all scanning windows that are divided into human body are made up, form final human detection result.
2. human body detecting method according to claim 1, wherein step (1) is described in the INRIA database, obtains negative sample by the bootstrapping operation, carries out as follows:
(2a) from the INRIA database, appoint first and get the positive sample of a part and negative sample;
(2b) extract respectively the HOG feature of positive negative sample, and based on edge feature and the textural characteristics of second generation curve wave conversion, and the AdaBoost sorter is carried out classification based training, obtain two preliminary classification devices;
(2c) use respectively this two preliminary classification devices, the non-human body image of in the test I NRIA database all the other, for these non-human body images, random choose part image and current negative sample form new negative sample collection in the scanning window image that is divided into human body image by mistake;
(2d) repeat (2b-2c), from the negative sample of gained, remove the negative sample that repeats, until reject the negative sample of all repetitions, obtain final negative sample collection.
3. human body detecting method according to claim 1, wherein the described extraction training sample of step (2) is concentrated the edge feature vector based on second generation curve wave conversion of each training sample, and the edge feature vector that extracts carried out the selection of optimum statistic, carry out as follows:
(3a) each training sample is done the second generation curve wave conversion of fast discrete, the parameter of second generation curve wave conversion is set to real-valued second generation curve wave conversion, and scale parameter is 3 layers, and inferior rough layer has 8 directional subbands, and detailed level is curve wave system number;
(3b) the identical sub-band coefficients matrix of size on inferior rough layer and the detailed level is spliced to form a large matrix of coefficients, and the matrix of coefficients of each yardstick carried out the subdivision of 8 * 8 pixels, with the fritter of each 8 * 8 pixel as a curve ripple piece;
(3c) calculate these the several statistics of energy, entropy, standard deviation, average, maximal value, minimum value and contrast of coefficient in each curve ripple piece, and the statistic of each curve ripple piece is carried out cascade, obtain initial edge feature vector; Extract the initial edge feature of all training samples, obtain a sample characteristics collection, and carry out classification based training with the AdaBoost algorithm;
Every kind of number of times that statistic is selected when (3d) statistics is trained the AdaBoost sorter, to screening in the statistic in (3c), the statistic that selected number of times is more wherein has as the statistical characteristic value of optimum: energy, entropy, contrast, standard deviation and maximal value;
(3e) optimum statistic with coefficient in all curve ripple pieces connects, and forms a proper vector, i.e. the edge feature vector.
4. human body detecting method according to claim 1, wherein the described extraction training sample of step (3) is concentrated the texture feature vector based on second generation curve wave conversion of each training sample, carries out as follows:
(4a) adopt 1/2 overlapping method to carry out the subdivision of 8 * 8 coefficient magnitude to the matrix of coefficients of the rough layer behind the second generation curve wave conversion of each training sample, and the coefficient fritter of each 8 * 8 size calculated its co-occurrence matrix, the quantification progression of this co-occurrence matrix is 16;
(4b) for the co-occurrence matrix of the coefficient fritter of each 8 * 8 size, calculate its angle second moment, entropy, contrast, correlativity, average and and variance and this six statistics, the statistic of the co-occurrence matrix of the coefficient fritter of all 8 * 8 sizes is connect, form proper vector, i.e. a texture feature vector.
5. human body detecting method according to claim 1, utilize characteristics of human body's vector of all training samples in step (2) to (4) the described method calculation training sample set, form the sample characteristics collection, and utilize sample characteristics set pair AdaBoost sorter to carry out classification based training, concrete steps are as follows;
(5a) will utilize the edge feature vector sum texture feature vector of each training sample of step (2) and (3) extraction to carry out cascade, can obtain characteristics of human body's vector of each training sample;
(5b) characteristics of human body's vector is the vector of a M dimension, it is N that training sample is concentrated the number of training sample, then the edge feature vector sum texture feature vector of all training sample extractions will form the matrix of a N * M dimension, as the input of AdaBoost algorithm, train a sorter.
6. human body detecting method according to claim 1, wherein the method for the sliding window scanning of the described employing of step (5) is carried out as follows to being scanned detection by altimetric image:
(6a) the input arbitrary size by altimetric image, be the proportional zoom of [0.4,0.5,0.6,0.7,0.8,0.9] with it by pantograph ratio, be the occurrence of the pantograph ratio that varies in size in the square bracket;
(6b) with behind the convergent-divergent by the altimetric image upper left corner and the zone size such as a training sample as first scanning window image, every to 8 pixels of right translation or downwards 8 pixels of translation obtain one group of scanning window as a new scanning window;
(6c) utilize respectively the method for step (2) and step (3) to calculate its edge feature vector sum texture feature vector to the image of each scanning window part;
The edge feature vector sum texture feature vector that (6d) scanning window image is partly extracted carries out characteristics of human body's vector that cascade obtains final scanning window image, and carries out the judgement of human body and non-human body with the resulting sorter of step (4).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201010259562 CN101930549B (en) | 2010-08-20 | 2010-08-20 | Second generation curvelet transform-based static human detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201010259562 CN101930549B (en) | 2010-08-20 | 2010-08-20 | Second generation curvelet transform-based static human detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101930549A CN101930549A (en) | 2010-12-29 |
| CN101930549B true CN101930549B (en) | 2013-02-27 |
Family
ID=43369714
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN 201010259562 Expired - Fee Related CN101930549B (en) | 2010-08-20 | 2010-08-20 | Second generation curvelet transform-based static human detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101930549B (en) |
Families Citing this family (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101984440A (en) * | 2010-10-26 | 2011-03-09 | 大连工业大学 | Holographic clothes design method |
| CN102426653B (en) * | 2011-10-28 | 2013-04-17 | 西安电子科技大学 | Static human body detection method based on second generation Bandelet transformation and star type model |
| CN102810159B (en) * | 2012-06-14 | 2014-10-29 | 西安电子科技大学 | Human body detecting method based on SURF (Speed Up Robust Feature) efficient matching kernel |
| CN102902950A (en) * | 2012-06-26 | 2013-01-30 | 西安电子科技大学 | Human movement identification method based on second-generation Curvelet transformation |
| CN103870795A (en) * | 2012-12-13 | 2014-06-18 | 北京捷成世纪科技股份有限公司 | Automatic detection method and device of video rolling subtitle |
| CN103942540A (en) * | 2014-04-10 | 2014-07-23 | 杭州景联文科技有限公司 | False fingerprint detection algorithm based on curvelet texture analysis and SVM-KNN classification |
| CN104091178A (en) * | 2014-07-01 | 2014-10-08 | 四川长虹电器股份有限公司 | Method for training human body sensing classifier based on HOG features |
| CN104090658A (en) * | 2014-07-04 | 2014-10-08 | 四川长虹电器股份有限公司 | Human body image detection method based on HOG features |
| CN106127233A (en) * | 2016-06-15 | 2016-11-16 | 天津中科智能识别产业技术研究院有限公司 | A kind of eyes open detection method and the system of closed state |
| CN107589449B (en) * | 2017-08-29 | 2020-04-28 | 电子科技大学 | Three-dimensional data fault enhancement method based on curve Gabor filtering |
| CN111214255B (en) * | 2020-01-12 | 2023-07-25 | 刘涛 | Medical ultrasonic image computer-aided method |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101430788A (en) * | 2008-11-19 | 2009-05-13 | 西安电子科技大学 | Lunette local Wiener filtering method based on second generation Curvelet transformation |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7840625B2 (en) * | 2005-04-07 | 2010-11-23 | California Institute Of Technology | Methods for performing fast discrete curvelet transforms of data |
-
2010
- 2010-08-20 CN CN 201010259562 patent/CN101930549B/en not_active Expired - Fee Related
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101430788A (en) * | 2008-11-19 | 2009-05-13 | 西安电子科技大学 | Lunette local Wiener filtering method based on second generation Curvelet transformation |
Non-Patent Citations (1)
| Title |
|---|
| 殷海青等.联合纹理提取和边缘检测的新方法.《系统工程与电子技术》.2010,第32卷(第4期), * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN101930549A (en) | 2010-12-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN101930549B (en) | Second generation curvelet transform-based static human detection method | |
| CN106874894B (en) | Human body target detection method based on regional full convolution neural network | |
| CN102682287B (en) | Pedestrian detection method based on saliency information | |
| CN103761531B (en) | The sparse coding license plate character recognition method of Shape-based interpolation contour feature | |
| CN104392463B (en) | Image salient region detection method based on joint sparse multi-scale fusion | |
| CN105825183B (en) | Facial expression recognizing method based on partial occlusion image | |
| CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
| CN102254196B (en) | Method for identifying handwritten Chinese character by virtue of computer | |
| CN105320950A (en) | A video human face living body detection method | |
| CN105160317A (en) | Pedestrian gender identification method based on regional blocks | |
| CN101859382A (en) | License plate detection and identification method based on maximum stable extremal region | |
| CN106778687A (en) | Method for viewing points detecting based on local evaluation and global optimization | |
| CN104778457A (en) | Video face identification algorithm on basis of multi-instance learning | |
| CN103902968A (en) | Pedestrian detection model training method based on AdaBoost classifier | |
| CN103186776B (en) | Based on the human body detecting method of multiple features and depth information | |
| CN103020614B (en) | Based on the human motion identification method that space-time interest points detects | |
| CN105205449A (en) | Sign language recognition method based on deep learning | |
| CN102867195B (en) | Method for detecting and identifying a plurality of types of objects in remote sensing image | |
| CN103971106A (en) | Multi-view human facial image gender identification method and device | |
| CN101520839B (en) | Human body detection method based on second-generation strip wave conversion | |
| CN103824062B (en) | Motion identification method for human body by parts based on non-negative matrix factorization | |
| CN102609715B (en) | Object type identification method combining plurality of interest point testers | |
| CN103455826A (en) | Efficient matching kernel body detection method based on rapid robustness characteristics | |
| CN103077383B (en) | Based on the human motion identification method of the Divisional of spatio-temporal gradient feature | |
| CN104268557A (en) | Polarization SAR classification method based on cooperative training and depth SVM |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20130227 Termination date: 20180820 |