CN104239896A - Method for classifying crowd density degrees in video image - Google Patents
Method for classifying crowd density degrees in video image Download PDFInfo
- Publication number
- CN104239896A CN104239896A CN201410447731.0A CN201410447731A CN104239896A CN 104239896 A CN104239896 A CN 104239896A CN 201410447731 A CN201410447731 A CN 201410447731A CN 104239896 A CN104239896 A CN 104239896A
- Authority
- CN
- China
- Prior art keywords
- sample
- classification
- image block
- sorter
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a method for classifying crowd density degrees in a video image. The method comprises the steps that firstly, an interest region is selected from the video image and is subjected to image block division and analysis according to a perspective model; secondly, multi-scale textural features are obtained for each image block; the video image is subjected to clustering analysis, a classifier model based on confidence degree analysis is established, multiple binary classifiers are optimized and combined by designing error correction output codes based on a binary-tree classification idea, confidence samples are extracted, and SVM binary classifiers are trained; finally, a channel transmission model is utilized to conduct decoding, and the crowd density degrees which the samples belong to are obtained according to the posterior probability maximum rule. According to the method, classification is conducted on the premise that sample sets and features are identical, the accuracy and generalization performance are both superior to those of a transmission classification model, and an idea is provided for solutions to various classification problems represented by crowd density estimation.
Description
Technical field
The present invention relates to the technical field of intelligent video analysis and computer vision, particularly a kind of crowd density in video image grade separation method.
Background technology
Crowd density hierarchical estimation is based on computer vision and mode identification technology, by carrying out analytical calculation to monitoring image or video, draws the quantization level of crowd density.Crowd density information is the strong foundation of population surveillance management on a large scale, and it can provide market the inner crowd density information by Different periods distribution to market or retail point, assist management laminated reason distribution services and management resource.It can be widely used in bus stop, passenger traffic and the Vomitory of the facility such as railway station, airport and the population surveillance of important area, obtain the accurate data of passenger number and distribution in real time, prevent the potential safety hazard caused because client is crowded, for scientific dispatch, safety guarantee provide foundation.
When crowd density is quantified as density rating, crowd density estimation is just converted into multicategory classification problem: input the characteristic quantity relevant to crowd, output density grade.The common method solving this multicategory classification problem has polynomial fitting method, neural network, recently clustering method etc.Funtcional relationship between polynomial fitting method often needs input and output makes certain hypothesis, and when hypothesis does not conform to the actual conditions, performance will be affected.Neural computing is more complicated, and to minimize experience error for target function, is easily absorbed in Local Minimum or produced study phenomenon.Nearest clustering method (K-MeansandNearestNeighbor, KMNN) be by the training sample of same density rating cluster in feature space, form M representative point (for different density ratings, M need not be identical), for the sample of a certain unknown classification, with the classification of its arest neighbors representative point or k neighbour's representative point for according to the classification determining this sample.This method is simple, but comparatively large for Selecting parameter dependence, when the improper meeting of Selecting parameter causes hydraulic performance decline, even produces over-fitting.
Statistical Learning Theory is that the best at present for small sample statistical estimate and prediction study is theoretical, it theoretically systematic study empirical risk minimization principle set up condition, under finite sample empiric risk and expected risk relation and how to utilize these theories to find the problems such as new learning principle and method.Support vector machine (SVM) has been widely used in a lot of fields of pattern-recognition and data mining as a kind of implementation method of Statistical Learning Theory.The essence of SVM classifier is linear two sorters, its structure based risk minimization is theoretical, in feature space, construction optimum segmentation lineoid, makes learner obtain global optimization, and meets certain upper bound in the expected risk of whole sample space with certain probability.
When utilizing SVM process multi-class problem, just need to construct suitable multi classifier.Based on this, the present invention proposes a kind of method of crowd density grade separation, and the svm classifier model based on Confidence Analysis solves the problems such as " neighbour is similar ", " smeared out boundary " of crowd density classification.At Chinese CNKI, incomparably, correlative theses and the patent of current Anomaly groups event detection are searched in the database such as Wei Pu, Chinese patent, CN101751553A, CN101587537A and CN102044073A propose a kind of method of crowd density analysis and prediction respectively, but there is following shortcoming: 1) cannot adapt to scene changes, model parameter needs repetitive learning; 2) higher at crowd density, when serious shielding, linear relationship is no longer set up, and causes error to increase; 3) method based on foreground detection can only be used for the stream of people that moves.
Summary of the invention
The technical problem to be solved in the present invention is, for the adjacent phase Sihe gradual change boundary feature of crowd density in video image grade separation, provides a kind of crowd density in video image grade separation method.
Crowd density in video image grade separation method of the present invention in turn includes the following steps:
S1: select an area-of-interest in video image, and stroke block analysis according to perspective model, it being carried out to image block;
S2: ask for multi-dimension texture feature for described each image block;
S3: carry out cluster analysis for video image sample, sets up the sorter model based on Confidence Analysis;
S4: obtain the crowd density grade belonging to image pattern according to the maximum rule of posterior probability.
In a kind of crowd density grade separation method of the present invention, described foundation is based on the sorter model of Confidence Analysis, and its step is as follows:
S1: the error correcting output code based on binary tree designs;
S2: the SVM based on confidence sample trains;
S3: based on transmission solution to model code calculation.
In foundation of the present invention based in the sorter model of Confidence Analysis, the error correcting output code design procedure based on binary tree is as follows:
S1: initialization, category set
, order
;
S2: establish
for one two sorting technique of current universal class, calculate two classification subsets of gained
with
between Fisher distance;
S3: for two classification arbitrary in set E
repeat Step2, ask two classification
, make two classification subsets of its correspondence
with
between Fisher distance be maximum;
S4: with
two sorters of representative are the row in error correcting output code matrix;
S5: make respectively
,
, and repeat Step1, Step2, Step3, until all only comprise a classification in all categories subset, export the error correcting output code of gained.
In foundation of the present invention based in the sorter model of Confidence Analysis, the SVM training step based on confidence sample is as follows:
S1: estimate posterior probability
;
S2: select confidence sample, weight is 1;
S3: utilize all confidence sample points (weight is not the training sample of 0), traditionally SVM training algorithm training SVM classifier.
In foundation of the present invention based in the sorter model of Confidence Analysis, as follows based on transmission solution to model code calculation:
In the mode of two sorters, its input end is the true classification of sample, 1 represents this sample should belong to positive classification in this sorter,-1 represents that this sample belongs to negative classification in this sorter, 0 represents that this sample does not relate in this sorter, its output terminal represents that this sample is through the sorted output classification of this sorter, 1 represents that this sample is divided into positive sample,-1 represents that this sample is divided into negative sample, the classification (classification 0) do not related to is divided in the middle of this two class with certain probability respectively, transformation parameter a in model, b, the available empirical value of c is estimated as follows:
，
，
，
Wherein, a represents the probability positive sample of input being divided into positive sample; B represent by input do not relate to the probability that classification sample is divided into positive sample; C represents probability input negative sample being divided into negative sample;
,
represent the sample class of input and output respectively;
,
represent
the input classification of individual sample and output classification;
represent the number of training sample;
represent indicative function,
。
In foundation of the present invention based in the sorter model of Confidence Analysis, the maximum rule of posterior probability is as follows:
Consider by
the categorizing system of individual two sorter compositions, supposes sample
output chain code be
, and classification
corresponding classification chain code is
, then maximum a posteriori probability
,
Wherein,
classification can be regarded as
prior probability, when can be set to without when priori
,
for classification number,
it is chain code
the probability occurred,
for likelihood function, suppose that each two sorters are separate, then have
Wherein,
,
,
be
the transformation parameter of individual sorter,
represent classification
in sample relative to
the input quantity of individual sorter, value is-1,0, or 1, represent such other style basis respectively in a jth sorter as negative sample, uncorrelated sample and positive sample;
be that two sorters export, value is-1 or 1,
value is 6 parameters in mode
one of, wherein mode parameter
can be estimated by the classification results of training sample set; At prior probability
when equal, maximize posterior probability and be converted into maximization likelihood function value, finally, sample
categorised decision be
.
In a kind of crowd density in video image grade separation method of the present invention, the described division based on perspective model, area-of-interest being carried out to image block, its step is as follows:
S1: the image block that the capacity that arranges is identical;
S2: the size of specifying smallest blocks and largest block image block;
S3: be similar to the size that perspective scale model obtains intermediate mass by linear difference.
In a kind of crowd density in video image grade separation method of the present invention, the described step asking for multi-dimension texture feature is as follows:
S1: video image sample is determined that the average division of yardstick obtains multilayer average image block;
S2: obtain image block Mean Matrix;
S3: determine n × n neighborhood submatrix, extracts the image block textural characteristics of individual layer block average;
S4: divide according to different scale multiple averaging the multilayer average image block that video image sample obtains different scale, take above-mentioned steps, calculate the image block textural characteristics obtaining different scale;
S5: combine the image texture block textural characteristics under described different scale, obtains the multi-dimension texture feature of image block.
Crowd density in video image grade separation method of the present invention, establish in different scene, different cameras angle, the diverse location hypograph block density rating criteria for classifying, under the prerequisite that sample set is identical with feature, classification accuracy rate and Generalization Capability are all better than traditional classification model, for being that the multicategory classification problem solving of representative provides a kind of thinking with crowd density estimation.
Accompanying drawing explanation
Fig. 1 is the process flow diagram of crowd density in video image grade separation method of the present invention;
Fig. 2 is the process flow diagram of the foundation in the present invention based on the sorter model of Confidence Analysis;
Fig. 3 is the error correcting output code design flow diagram based on binary tree in the present invention;
Fig. 4 is that the SVM based on confidence sample in the present invention trains process flow diagram;
Fig. 5 be in the present invention based on transmission solution to model code process flow diagram.
Embodiment
To be described in detail to the present invention below in conjunction with drawings and Examples, to carry out more deep understanding to object of the present invention, feature and advantage.
Basic scheme of the present invention is, carries out image block division based on perspective model to target video image region, and texture feature extraction in units of video image blocks also carries out cluster analysis; By the error correcting output codes of design based on binary tree sort thought, multiple two sorters of optimal combination, set up the sorter model based on Confidence Analysis; Then extract confidence sample, training SVM bis-sorter, utilizes transmission model to decode; Finally, the crowd density grade belonging to sample is obtained according to the maximum rule of posterior probability.
Fig. 1 is the process flow diagram of crowd density in video image grade separation method of the present invention, specifically can comprise the following steps:
S1: first select an interested region in target video image 100, and stroke block 102 according to perspective model 101, it being carried out to image block;
The present embodiment determines the relation between image block Local textural feature and texture level based on perspective model 101.Can produce visual angle distortion when utilizing two dimensional image to represent three-dimensional spatial information, this three-dimensional is similar to the perspective model 101 of two dimension with linear transformation:
, wherein, H is perspective transformation matrix.The Exact Solution of H needs the information using camera calibration, in practical operation, can by state specify the most nearby with the block size of farthest, then interpolation is carried out to intermediate mass and obtains block size.By the correction of being out of shape visual angle, ensure that the actual area area of Same Scene hypograph block representative is substantially identical.
To needing the region of carrying out crowd density estimation, arranged by man-machine interaction and be divided into image block, and all image blocks are demarcated as 5 grades according to crowd density, the crowd density represented respectively in block is very low (1), low (2), medium (3), high (4) and very high (5), the criteria for classifying of image block density rating is mainly according to the number comprised in image block and the area ratio occupied in image block of crowd, and classification thresholds refers to table 1.Table 1 provides the demarcation embodiment to outdoor certain square crowd, area-of-interest is set by hand in figure, according to camera parameters and perspective model, be divided into the image block of " near big and far smaller ", allow to there is certain overlapping region between adjacent image block, to cover area-of-interest completely as far as possible, each image block is demarcated as 5 density ratings (1 ~ 5) according to its crowd massing degree, after the density rating determining each image block, the total body density grade of area-of-interest:
, wherein, N represents the image block number in area-of-interest,
represent the density rating of i-th image block,
for total body density grade,
it is right to represent
employing rounding-off method rounds.
One embodiment form of the table 1 image block density rating criteria for classifying;
| Density rating | Very low (1) | Low (2) | In (3) | High (4) | Very high (5) |
| Number | 0-0.5 | 0.5-2 | 2-4 | 4-6 | >6 |
| Crowd's area occupied | <10% | 10%-30% | 30%-60% | 60%-90% | >90% |
S2: ask for multi-dimension texture feature 103 for described each image block;
The local binarization operator (Local Binary Pattern, LBP) that the present invention improves describes crowd's local grain, and adds up its multiple dimensioned histogram on image block as image block characteristics.LBP mode computation formula is expressed as follows:
, wherein,
represent central element gray-scale value,
represent neighborhood element gray-scale value.
S3: carry out cluster analysis for video image sample, sets up the sorter model 104 based on Confidence Analysis;
Cluster analysis in the textural feature space of this image block, carries out K mean cluster to the image block of video image under same density rating, and obtain K cluster, K cluster just represents the distribution form of this density rating in this textural feature space.Different density ratings, K value is not necessarily identical, and K refers to cluster number.The cluster centre that cluster uses and cluster radius use following formulae discovery respectively:
Cluster centre:
Cluster radius:
;
S4: obtain the crowd density grade 105 belonging to sample according to the maximum rule of posterior probability.
Consider by
the categorizing system of individual two sorter compositions, supposes sample
output chain code be
, and classification
corresponding classification chain code is
, then maximum a posteriori probability
,
Wherein,
classification can be regarded as
prior probability, when can be set to without when priori
,
for classification number,
it is chain code
the probability occurred,
for likelihood function, suppose that each two sorters are separate, then have
Wherein,
,
,
be
the transformation parameter of individual sorter,
represent classification
in sample relative to
the input quantity of individual sorter, value is-1,0, or 1, represent such other style basis respectively in a jth sorter as negative sample, uncorrelated sample and positive sample;
be that two sorters export, value is-1 or 1,
value is 6 parameters in mode
one of, wherein mode parameter
can be estimated by the classification results of training sample set; At prior probability
when equal, maximize posterior probability and be converted into maximization likelihood function value, finally, sample
categorised decision be
.
Fig. 2 is the process flow diagram of the foundation in the present invention based on the sorter model of Confidence Analysis, and its step is as follows:
S1: based on the error correcting output code design 100 of binary tree;
S2: based on the SVM training 101 of confidence sample;
S3: based on transmission solution to model code calculation 102.
Fig. 3 is the error correcting output code design flow diagram based on binary tree in the present invention, and its step is as follows:
S1: initialization 200, category set
, order
;
S2: establish
for current universal class
one two sorting technique, calculate two classification subsets of gained
with
between Fisher distance 201:
Wherein,
with
represent two sons respectively and concentrate the average comprising sample,
with
represent the variance comprising sample in two subsets respectively;
S3: for two classification arbitrary in set E
repeat Step2, ask two classification
, make two classification subsets of its correspondence
with
between Fisher distance be maximum 202, namely
S4: with
two sorters of representative are the row 203 in error correcting output code matrix;
S5: make respectively
,
, and repeat Step1, Step2, Step3, until all only comprise a classification 204 in all categories subset, export the error correcting output code 205 of gained.
Fig. 4 is that the SVM based on confidence sample in the present invention trains process flow diagram, and its step is as follows:
S1: posterior probability
estimation 300, adopt
nearest neighbour method, gets
near
individual sample, statistics wherein belongs to counting of positive classification
, then
;
S2: confidence samples selection 301, for threshold value
(
), if
, then
for non-confidence sample, weight is 0; Otherwise be confidence sample, weight is 1;
S3: utilize all confidence sample points (weight is not the training sample of 0), traditionally SVM training algorithm training SVM classifier 302.
Fig. 5 be in the present invention based on transmission solution to model code process flow diagram, its step is as follows:
In the mode of two sorters, its input end is the true classification of sample, 1 represents this sample should belong to positive classification in this sorter,-1 represents that this sample belongs to negative classification in this sorter, 0 represents that this sample does not relate in this sorter, its output terminal represents that this sample is through this sorter sorted output classification: 1 represents that this sample is divided into positive sample,-1 represents that this sample is divided into negative sample, the classification (classification 0) do not related to is divided in the middle of this two class with certain probability respectively, transformation parameter a in model, b, the available empirical value of c is estimated as follows:
，
，
，
Wherein, a represents the probability positive sample of input being divided into positive sample; B represent by input do not relate to the probability that classification sample is divided into positive sample; C represents probability input negative sample being divided into negative sample;
,
represent the sample class of input and output respectively;
,
represent
the input classification of individual sample and output classification;
represent the number of training sample;
represent indicative function,
.
Claims (8)
1. a crowd density in video image grade separation method, is characterized in that, comprises the steps:
S1: select an area-of-interest in video image, and stroke block analysis according to perspective model, it being carried out to image block;
S2: ask for multi-dimension texture feature for described each image block;
S3: carry out cluster analysis for video image sample, sets up the sorter model based on Confidence Analysis;
S4: obtain the crowd density grade belonging to image pattern according to the maximum rule of posterior probability.
2. method according to claim 1, is characterized in that, the sorter model set up based on Confidence Analysis comprises the steps:
S1: the error correcting output code based on binary tree designs;
S2: the SVM based on confidence sample trains;
S3: based on transmission solution to model code calculation.
3. method according to claim 2, is characterized in that, the error correcting output code design based on binary tree comprises the steps:
S1: initialization, category set
, order
;
S2: establish
for current universal class
one two sorting technique, calculate two classification subsets of gained
with
between Fisher distance;
S3: for two classification arbitrary in set E
repeat Step2, ask two classification
, make two classification subsets of its correspondence
with
between Fisher distance be maximum;
S4: with
two sorters of representative are the row in error correcting output code matrix;
S5: make respectively
,
, and repeat Step1, Step2, Step3, until all only comprise a classification in all categories subset, export the error correcting output code of gained.
4. method according to claim 2, is characterized in that, the SVM training based on confidence sample comprises the steps:
S1: estimate posterior probability
;
S2: select confidence sample, weight is 1;
S3: utilize all confidence sample points, traditionally SVM training algorithm training SVM classifier.
5. method according to claim 2, is characterized in that, as follows based on transmission solution to model code calculation content:
In the mode of two sorters, its input end is the true classification of sample, 1 represents this sample should belong to positive classification in this sorter,-1 represents that this sample belongs to negative classification in this sorter, 0 represents that this sample does not relate in this sorter, its output terminal represents that this sample is through the sorted output classification of this sorter, 1 represents that this sample is divided into positive sample,-1 represents that this sample is divided into negative sample, the classification do not related to is divided in the middle of this two class with certain probability respectively, and the available empirical value of transformation parameter a, b, c in model is estimated as follows:
，
，
，
Wherein, a represents the probability positive sample of input being divided into positive sample; B represent by input do not relate to the probability that classification sample is divided into positive sample; C represents probability input negative sample being divided into negative sample;
,
represent the sample class of input and output respectively;
,
represent
the input classification of individual sample and output classification;
represent the number of training sample;
represent indicative function,
。
6. method according to claim 1, is characterized in that, posterior probability maximum rule content is as follows:
Consider by
the categorizing system of individual two sorter compositions, supposes sample
output chain code be
, and classification
corresponding classification chain code is
, then maximum a posteriori probability
,
Wherein,
classification can be regarded as
prior probability, when can be set to without when priori
,
for classification number,
it is chain code
the probability occurred,
for likelihood function, suppose that each two sorters are separate, then have
Wherein,
,
,
be
the transformation parameter of individual sorter,
represent classification
in sample relative to
the input quantity of individual sorter, value is-1,0, or 1, represent such other style basis respectively in a jth sorter as negative sample, uncorrelated sample and positive sample;
be that two sorters export, value is-1 or 1,
value is 6 parameters in mode
one of, wherein mode parameter
can be estimated by the classification results of training sample set; At prior probability
when equal, maximize posterior probability and be converted into maximization likelihood function value, finally, sample
categorised decision be
.
7. method according to claim 1, is characterized in that, the partiting step carrying out image block to area-of-interest based on perspective model is as follows:
S1: the image block that the capacity that arranges is identical;
S2: the size of specifying smallest blocks and largest block image block;
S3: be similar to the size that perspective scale model obtains intermediate mass by linear difference.
8. method according to claim 1, is characterized in that, the described step asking for multi-dimension texture feature is as follows:
S1: video image sample is determined that the average division of yardstick obtains multilayer average image block;
S2: obtain image block Mean Matrix;
S3: determine n × n neighborhood submatrix, extracts the image block textural characteristics of individual layer block average;
S4: divide according to different scale multiple averaging the multilayer average image block that video image sample obtains different scale, take above-mentioned steps, calculate the image block textural characteristics obtaining different scale;
S5: combine the image texture block textural characteristics under described different scale, obtains the multi-dimension texture feature of image block.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410447731.0A CN104239896A (en) | 2014-09-04 | 2014-09-04 | Method for classifying crowd density degrees in video image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410447731.0A CN104239896A (en) | 2014-09-04 | 2014-09-04 | Method for classifying crowd density degrees in video image |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN104239896A true CN104239896A (en) | 2014-12-24 |
Family
ID=52227919
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410447731.0A Pending CN104239896A (en) | 2014-09-04 | 2014-09-04 | Method for classifying crowd density degrees in video image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104239896A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105631858A (en) * | 2015-12-21 | 2016-06-01 | 北京大学深圳研究生院 | Sample block-based image target counting method |
| CN107657274A (en) * | 2017-09-20 | 2018-02-02 | 浙江大学 | A kind of y-bend SVM tree unbalanced data industry Fault Classifications based on k means |
| CN107862230A (en) * | 2017-11-14 | 2018-03-30 | 安徽虚境科技有限公司 | A kind of grade determines and input method |
| CN107967688A (en) * | 2017-12-21 | 2018-04-27 | 联想(北京)有限公司 | The method and system split to the object in image |
| CN109829503A (en) * | 2019-02-12 | 2019-05-31 | 厦门美柚信息科技有限公司 | A kind of intensive frightened picture method of discrimination, system, equipment and its storage medium |
| CN111652257A (en) * | 2019-03-27 | 2020-09-11 | 上海铼锶信息技术有限公司 | Sample data cleaning method and system |
| CN111681763A (en) * | 2020-04-16 | 2020-09-18 | 北京大学 | Total knee replacement prosthesis model prediction method based on error correction output codes and electronic device |
| CN112184523A (en) * | 2020-09-27 | 2021-01-05 | 中南大学 | Subway carriage passenger guidance method and system based on environmental monitoring and illumination guidance |
| WO2021023310A1 (en) * | 2019-08-07 | 2021-02-11 | 中兴通讯股份有限公司 | Method for estimating and presenting passenger flow, system, and computer readable storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090316986A1 (en) * | 2008-04-25 | 2009-12-24 | Microsoft Corporation | Feature selection and extraction |
| CN101727573A (en) * | 2008-10-13 | 2010-06-09 | 汉王科技股份有限公司 | Method and device for estimating crowd density in video image |
| CN102044073A (en) * | 2009-10-09 | 2011-05-04 | 汉王科技股份有限公司 | Method and system for judging crowd density in image |
| CN103971123A (en) * | 2014-05-04 | 2014-08-06 | 南京师范大学 | Hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL) |
-
2014
- 2014-09-04 CN CN201410447731.0A patent/CN104239896A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090316986A1 (en) * | 2008-04-25 | 2009-12-24 | Microsoft Corporation | Feature selection and extraction |
| CN101727573A (en) * | 2008-10-13 | 2010-06-09 | 汉王科技股份有限公司 | Method and device for estimating crowd density in video image |
| CN102044073A (en) * | 2009-10-09 | 2011-05-04 | 汉王科技股份有限公司 | Method and system for judging crowd density in image |
| CN103971123A (en) * | 2014-05-04 | 2014-08-06 | 南京师范大学 | Hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105631858B (en) * | 2015-12-21 | 2018-06-22 | 北京大学深圳研究生院 | Image object method of counting based on sample block |
| CN105631858A (en) * | 2015-12-21 | 2016-06-01 | 北京大学深圳研究生院 | Sample block-based image target counting method |
| CN107657274A (en) * | 2017-09-20 | 2018-02-02 | 浙江大学 | A kind of y-bend SVM tree unbalanced data industry Fault Classifications based on k means |
| CN107862230A (en) * | 2017-11-14 | 2018-03-30 | 安徽虚境科技有限公司 | A kind of grade determines and input method |
| CN107967688A (en) * | 2017-12-21 | 2018-04-27 | 联想(北京)有限公司 | The method and system split to the object in image |
| CN109829503B (en) * | 2019-02-12 | 2021-12-17 | 厦门美柚股份有限公司 | Dense fear picture distinguishing method, system, equipment and storage medium thereof |
| CN109829503A (en) * | 2019-02-12 | 2019-05-31 | 厦门美柚信息科技有限公司 | A kind of intensive frightened picture method of discrimination, system, equipment and its storage medium |
| CN111652257A (en) * | 2019-03-27 | 2020-09-11 | 上海铼锶信息技术有限公司 | Sample data cleaning method and system |
| WO2021023310A1 (en) * | 2019-08-07 | 2021-02-11 | 中兴通讯股份有限公司 | Method for estimating and presenting passenger flow, system, and computer readable storage medium |
| CN111681763A (en) * | 2020-04-16 | 2020-09-18 | 北京大学 | Total knee replacement prosthesis model prediction method based on error correction output codes and electronic device |
| CN111681763B (en) * | 2020-04-16 | 2023-01-17 | 北京大学 | Total knee arthroplasty prosthesis model prediction method based on error correction output code and electronic device |
| CN112184523A (en) * | 2020-09-27 | 2021-01-05 | 中南大学 | Subway carriage passenger guidance method and system based on environmental monitoring and illumination guidance |
| CN112184523B (en) * | 2020-09-27 | 2022-06-07 | 中南大学 | Subway carriage passenger guidance method and system based on environmental monitoring and illumination guidance |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104239896A (en) | Method for classifying crowd density degrees in video image | |
| CN102044073B (en) | Method and system for judging crowd density in image | |
| CN109344736B (en) | Static image crowd counting method based on joint learning | |
| CN103839065B (en) | Extraction method for dynamic crowd gathering characteristics | |
| CN107657226B (en) | People number estimation method based on deep learning | |
| CN110059581A (en) | People counting method based on depth information of scene | |
| CN101727573B (en) | Method and device for estimating crowd density in video image | |
| CN107633226B (en) | Human body motion tracking feature processing method | |
| CN109543695A (en) | General density people counting method based on multiple dimensioned deep learning | |
| CN101329734A (en) | License plate character recognition method based on K-L transform and LS-SVM | |
| CN104182772A (en) | Gesture recognition method based on deep learning | |
| CN104134068B (en) | Monitoring vehicle feature representation and classification method based on sparse coding | |
| CN102663100A (en) | Two-stage hybrid particle swarm optimization clustering method | |
| CN104751136A (en) | Face recognition based multi-camera video event retrospective trace method | |
| CN102254006A (en) | Method for retrieving Internet video based on contents | |
| CN107958260B (en) | Group behavior analysis method based on multi-feature fusion | |
| CN108154158B (en) | Building image segmentation method for augmented reality application | |
| CN103914705A (en) | Hyperspectral image classification and wave band selection method based on multi-target immune cloning | |
| CN102930294A (en) | Chaotic characteristic parameter-based motion mode video segmentation and traffic condition identification method | |
| CN107609509A (en) | A kind of action identification method based on motion salient region detection | |
| CN115527133A (en) | High-resolution image background optimization method based on target density information | |
| Ji et al. | Semisupervised hyperspectral image classification using spatial-spectral information and landscape features | |
| Bisio et al. | Traffic analysis through deep-learning-based image segmentation from UAV streaming | |
| CN114359632A (en) | Point cloud target classification method based on improved PointNet + + neural network | |
| CN104732246B (en) | A kind of semi-supervised coorinated training hyperspectral image classification method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20141224 |