CN104616032B - Multi-camera system target matching method based on depth convolutional neural networks - Google Patents
Multi-camera system target matching method based on depth convolutional neural networks Download PDFInfo
- Publication number
- CN104616032B CN104616032B CN201510047118.4A CN201510047118A CN104616032B CN 104616032 B CN104616032 B CN 104616032B CN 201510047118 A CN201510047118 A CN 201510047118A CN 104616032 B CN104616032 B CN 104616032B
- Authority
- CN
- China
- Prior art keywords
- image
- target
- ith
- representing
- block
- 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
- 238000000034 method Methods 0.000 title claims abstract description 23
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 13
- 238000005070 sampling Methods 0.000 claims abstract description 21
- 238000012706 support-vector machine Methods 0.000 claims abstract description 8
- 238000000605 extraction Methods 0.000 claims abstract description 6
- 239000013598 vector Substances 0.000 claims description 27
- 239000011159 matrix material Substances 0.000 claims description 21
- 238000012545 processing Methods 0.000 claims description 12
- 238000013507 mapping Methods 0.000 claims description 8
- 238000011176 pooling Methods 0.000 claims description 6
- 238000010276 construction Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000013139 quantization Methods 0.000 claims description 3
- 238000012549 training Methods 0.000 claims description 3
- 238000012544 monitoring process Methods 0.000 abstract description 6
- 230000006870 function Effects 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 2
- 230000001131 transforming effect Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000001186 cumulative effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
A kind of target matching method between multiple-camera based on depth convolutional neural networks.The present invention initializes multiple convolution kernels based on localised protection projecting method, and down-sampling is carried out to image based on maximum pond method, passes through successively eigentransformation, extraction more robust, more representative histogram feature；Multi-class support vector machine SVM classifier is recycled to carry out Classification and Identification.When target enters another camera coverage domain from a camera coverage domain, it is extracted feature and mark corresponding to target labels, realize multiple-camera cooperation monitoring field target is accurately identified, for target handoff and tracking etc..
Description
Technical Field
The invention belongs to the field of intelligent video monitoring in computer vision, and discloses a target matching method based on a deep convolutional neural network, which is suitable for a multi-camera cooperative video monitoring system.
Background
In a large-scale video monitoring place, such as an airport, a subway station, a square and the like, a target in a multi-camera cooperation monitoring system is tracked, and target matching among multiple cameras is a key step. For a large video monitoring scene, calibration of cameras is difficult and complex, and the spatial relationship, the temporal relationship and the time difference among the cameras are difficult to infer, so that the current widely-applied target matching method among multiple cameras is mainly based on target matching of features, and the accuracy of a matching result is directly influenced by the effectiveness of feature selection. But the extraction of robust features that can effectively characterize the target is a difficult problem. The features commonly used at present include color, texture, etc., and the features are difficult to maintain good robustness in all monitored scenes. Therefore, we propose a target matching method based on deep learning, which can adaptively learn features from a video frame sequence to achieve accurate target matching. Compared with the traditional neural network, the deep neural network overcomes the problem of less network layers, obtains more abstract feature expression by performing layer-by-layer transformation on the features, realizes the target classification as the final output layer of the network, and greatly improves the speed and efficiency of target matching.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provides a target matching method among multiple cameras based on deep learning, which comprises the following specific steps:
(1) preprocessing of the target image: extracting n target images of a multi-camera domain, and dividing the target images into m labels; uniformly adjusting the size of the image to h multiplied by w by utilizing a bicubic interpolation algorithm (bicubic interpolation), wherein h is the height of the image, and w is the width of the image; simply scaling the pixel values of the image samples so that the final pixel values all fall between [0,1 ]; the labels of the n images are stored as n x 1 data, and the value range of each label is [1, …, m ];
(2) extracting features based on a deep convolutional neural network:
(a) selecting n from the target image extracted in the step (1)tA training sample as a volumePerception node of first-layer input layer of cumulative neural networkWherein, Xi，i＝1,2,…,ntRepresenting the ith image;
(b) the filter applied to target image feature extraction is a convolution kernel constructed based on a local preserving projection method, and the specific construction method is as follows:
for image XiPerforming block processing to set block size to p1×p2Then XiAll the blocks of (a) are:wherein x isi,jJ is 1, …, hw represents XiThe jth block vector of (1); the block mean is then subtracted from each block to yield: wherein,j is 1, …, hw denotes the block after mean removal; the same processing is done for all input images X, resulting in:
the feature vector is calculated as follows: XLXTa＝λXDXTa, wherein a is an eigenvector, λ is an eigenvalue corresponding to a, and D is a diagonal matrix, and element values of the diagonal matrix are columns and/or rows of the weight matrix W; the weight matrix W is of dimension nt×ntOf a sparse matrix, WijRepresenting a sampleAndthe Euclidean distance between all samples is calculated, and for each sample, the k closest to the sample is foundnearestOne sample, namely: if the sampleIn a sampleK of (a)nearestWithin a nearest neighbor, or sampleIn a sampleK of (a)nearestWithin a nearest neighbor, thenOtherwise, Wij＝0；Dii＝∑jWjiAnd L-D is a Laplacian matrix; sorting the calculated eigenvectors according to the magnitude of the eigenvalues, and taking the top k1A feature vectorLet Vi 1＝ai-1，i＝1,2,…,k1Then, thenNamely, the extracted convolution kernel is obtained;
a convolution kernel V1With each frame imageCarry out convolution, i.e.i＝1,…,nt，j＝1,…,k1Then n is generated in the convolutional layertk1A web output feature map represented as:
(c) performing feature point down-sampling on the obtained feature map Y based on maximum value pooling (max boosting); let the sampling window size be s1×s1Then n is obtainedtk1Amplitude output feature map:
wherein, the ith output characteristic mapping chartJ row and k column of the ith output characteristic mapping chart, i is 1, …, ntk1， u, v denote the sampling step size, YiRepresenting the ith input image, and max { } representing a maximum function; in addition, the algorithm uses non-overlapping sampling, i.e., taking u-v-s1；
(d) Using a procedure similar to (b) for image XiPerforming block processing to set block size to p1×p2Taking the feature map Z obtained in the step (c) as the input of the convolutional layer, and removing the mean value of the block data of each frame of image to obtain an input image:
wherein, the ith input feature mapi＝1,…,ntk1Representing the ith block-de-averaged image,j is 1, …, hw represents j block vector of ith image after mean value is removed; constructing a weight matrix W and according to ZLZTa＝λZDZTa, calculating eigenvectors, sorting according to the magnitude of eigenvalues, and taking the top k2The feature vector is used as the selected convolution kernelWherein, Vi 2，i＝1,…,k2Represents V2The ith convolution kernel of (a); the resulting convolution kernel V is then used2For each frame of imagePerforming convolution to generate n on the convolution layertk1k2Amplitude output feature map:
wherein,i＝1,…,ntk1，j＝1,…,k2；
(e) performing feature point down-sampling on the obtained feature map U based on a maximum value pooling method by adopting the steps similar to the step (c); let the sampling window size be s2×s2Then n is obtainedtk1k2Amplitude output feature map:
wherein, the ith output characteristic mapping chart J row and k column of the ith output characteristic mapping chart, i is 1, …, ntk1k2，U, v denote the sampling step size, UiRepresenting the ith input image, and max { } representing a maximum function; in addition, the algorithm uses non-overlapping sampling, i.e., taking u-v-s2；
(f) Order toi＝1,…,ntk1I.e. take every k in O2The images are grouped, and are subjected to Heaviside binary quantization and then are processed into decimal values, and each k is2Converting a picture into an imagei＝1,…,ntk1Wherein H (. cndot.) represents the Heaviside function, Pi jRepresents PiJ-th image of (1), TiRepresenting a decimal processing result, having a value range ofThen every k is taken1Breadth TiDividing each image into a group, dividing each image into B blocks, calculating histogram features of each block region, connecting the histogram features of the B blocks into row vectors, and defining the row vectors asWherein,l＝1,…,nt，s＝1,…,k1(ii) a Then for each image X in (a)lFinally extracting the feature vector based on the convolutional neural networkl＝1,…,nt；
(3) Classification and identification: extracting the above-mentioned featuresAnd as input, the target label corresponding to each feature vector is used as output, and a classifier model of the target is constructed by a multi-class Support Vector Machine (SVM). Based on the classifier model, the targets in different camera fields can be labeled and classified for target handover, tracking and the like.
The invention has the beneficial effects that:
the invention adopts a local preserving projection method to initialize the convolution kernel, but not to randomly initialize the convolution kernel, so that the obvious characteristics of the target image can be accurately preserved, thereby leading the extracted histogram characteristics to be capable of keeping the invariance to the scale change and the rotation of the target, having stronger adaptability to the illumination change of a scene and greatly improving the identification rate of the target. The method performs downsampling on the image convolved by using the local preserving projection method, effectively reduces the feature dimension, avoids dimension disaster, greatly shortens the identification time of the target, and effectively eliminates the reduction of the identification rate caused by dimension reduction by adopting multi-convolution to perform convolution on the image and overlapping the features.
Drawings
FIG. 1 is a flow chart of the present invention.
Detailed Description
The method comprises two parts of target feature extraction and target classification identification. The method comprises the steps of extracting the characteristics of a target, constructing a neural network with multiple hidden layers, transforming the characteristics of the sample layer by layer, transforming the characteristic representation of the sample in an original space to a new characteristic space to learn more useful characteristics, and then using the characteristics as input characteristics of a multi-class SVM classifier to perform classification and identification on the target, thereby finally improving the accuracy of classification or prediction. Fig. 1 shows an implementation block diagram of the algorithm, and the specific steps are as follows:
(1) preprocessing of the target image: extracting n target images of a multi-camera domain, and dividing the target images into m labels; uniformly adjusting the size of the image to h multiplied by w by utilizing a bicubic interpolation algorithm (bicubic interpolation), wherein h is the height of the image, and w is the width of the image; simply scaling the pixel values of the image samples so that the final pixel values all fall between [0,1 ]; the labels of the n images are stored as n x 1 data, and the value range of each label is [1, …, m ]; 001 represents the preprocessed image;
(2) extracting features based on a deep convolutional neural network:
(a) select n from 001tTraining samples as sensing nodes of the first input layer of the convolutional neural networkNamely 002, wherein, Xi，i＝1,2,…,ntRepresenting the ith image;
(b) the filter applied to target image feature extraction is a convolution kernel constructed based on a local preserving projection method, and the specific construction method is as follows:
for image X in 002iPerforming block processing to set block size to p1×p2Then XiAll the blocks of (a) are:wherein x isi,jJ is 1, …, hw represents XiThe jth block vector of (1); the block mean is then subtracted from each block to yield:wherein,j is 1, …, hw denotes the block after mean removal; the same processing is done for all input images X, resulting in: namely 003;
the feature vector is calculated as follows: XLXTa＝λXDXTa, wherein a is an eigenvector, λ is an eigenvalue corresponding to a, and D is a diagonal matrix, and element values of the diagonal matrix are columns and/or rows of the weight matrix W; the weight matrix W is of dimension nt×ntOf a sparse matrix, WijRepresenting a sampleAndthe Euclidean distance between all samples is calculated, and for each sample, the k closest to the sample is foundnearestOne sample, namely: if the sampleIn a sampleK of (a)nearestWithin a nearest neighbor, or sampleIn a sampleK of (a)nearestWithin a nearest neighbor, thenOtherwise, Wij＝0；Dii＝∑jWjiAnd L-D is a Laplacian matrix; sorting the calculated eigenvectors according to the magnitude of the eigenvalues, and taking the top k1A feature vectorLet Vi 1＝ai-1，i＝1,2,…,k1Then, thenNamely, the extracted convolution kernel is obtained;
a convolution kernel V1And 003 in each frame imagePerforming convolution to obtaini＝1,…,nt，j＝1,…,k1Then n is generated in the convolutional layertk1A web output feature map represented as:
namely 004;
(c) for the feature map Y shown in 004, feature point downsampling is performed on the basis of maximum value pooling (max boosting); let the sampling window size be s1×s1Then n is obtainedtk1Amplitude output feature map:
005, wherein the ith output characteristic map J row and k column of the ith output characteristic mapping chart, i is 1, …, ntk1，u, v denote the sampling step size, YiRepresenting the ith input image, and max { } representing a maximum function; in addition, the algorithm uses non-overlapping sampling, i.e., taking u-v-s1；
(d) Using a procedure similar to (b) for image X in 005iPerforming block processing to set block size to p1×p2Taking the feature map Z obtained in the step (c) as the input of the convolutional layer, and removing the mean value of the block data of each frame of image to obtain an input image:
i.e., 006, wherein the ith input feature mapi＝1,…,ntk1Representing the ith block-de-averaged image,j is 1, …, hw represents j block vector of ith image after mean value is removed; constructing a weight matrix W and according to ZLZTa＝λZDZTa, calculating eigenvectors, sorting according to the magnitude of eigenvalues, and taking the top k2The feature vector is used as the selected convolution kernelWherein, Vi 2，i＝1,…,k2Represents V2The ith convolution kernel of (a); the resulting convolution kernel V is then used2For each frame of imagePerforming convolution to generate n on the convolution layertk1k2Amplitude output feature map:
i.e., 007, wherein,i＝1,…,ntk1，j＝1,…,k2；
(e) for the feature map U shown by 007, the feature points are downsampled based on the maximum value pooling by the steps similar to (c); let the sampling window size be s2×s2Then n is obtainedtk1k2Amplitude output feature map:
i.e., 008, where the ith output feature map J row and k column of the ith output characteristic mapping chart, i is 1, …, ntk1k2，U, v denote the sampling step size, UiRepresenting the ith input image, and max { } representing a maximum function; in addition, the algorithm uses non-overlapping sampling, i.e., taking u-v-s2；
(f) Order toi＝1,…,ntk1Taking each k in O as shown in 0082The images are grouped, and are subjected to Heaviside binary quantization and then are processed into decimal values, and each k is2Converting a picture into an imagei＝1,…,ntk1Wherein H (. cndot.) represents the Heaviside function, Pi jRepresents PiJ-th image of (1), TiRepresenting a decimal processing result, having a value range ofThen every k is taken1Breadth TiDividing each image into a group, dividing each image into B blocks, calculating histogram features of each block region, connecting the histogram features of the B blocks into row vectors, and defining the row vectors asWherein,l＝1,…,nt，s＝1,…,k1(ii) a Then for each image X in 002lFinally extracting the feature vector based on the convolutional neural networkl＝1,…,nt；
(3) Classification and identification: extracting the above-mentioned featuresAnd as input, the target label corresponding to each feature vector is used as output, and a classifier model of the target is constructed by a multi-class Support Vector Machine (SVM). Based on the model of the classifier, it is,the method can realize labeling and classification of targets in different camera fields, and is used for target handover, tracking and the like.
The embodiments described in this specification are merely illustrative of implementations of the inventive concept and the scope of the present invention should not be considered limited to the specific forms set forth in the embodiments but rather by the equivalents thereof as may occur to those skilled in the art upon consideration of the present inventive concept.
Claims (1)
1. A multi-camera system target matching method based on a deep convolutional neural network is characterized by comprising the following steps:
(1) preprocessing of the target image: extracting n target images of a multi-camera domain, and dividing the target images into m labels; uniformly adjusting the size of the image to h multiplied by w by utilizing a bicubic interpolation algorithm (bicubic interpolation), wherein h is the height of the image, and w is the width of the image; simply scaling the pixel values of the image samples so that the final pixel values all fall between [0,1 ]; the labels of the n images are stored as n x 1 data, and the value range of each label is [1, …, m ];
(2) extracting features based on a deep convolutional neural network:
(a) selecting n from the target image extracted in the step (1)tTraining samples as sensing nodes of the first input layer of the convolutional neural networkWherein, Xi，i＝1,2,…,ntRepresenting the ith image;
(b) the filter applied to target image feature extraction is a convolution kernel constructed based on a local preserving projection method, and the specific construction method is as follows:
for image XiPerforming block processing to set block size to p1×p2Then XiAll the blocks of (a) are:wherein x isi,jJ is 1, …, hw represents XiThe jth block vector of (1); the block mean is then subtracted from each block to yield: wherein,representing the blocks with the mean removed; the same processing is done for all input images X, resulting in:
the feature vector is calculated as follows: XLXTa＝λXDXTa, where a is the feature vector and λ is aCorresponding characteristic value D is diagonal matrix, and element values of the diagonal matrix are columns and/or rows of the weight matrix W; the weight matrix W is of dimension nt×ntOf a sparse matrix, WijRepresenting a sampleAndthe Euclidean distance between all samples is calculated, and for each sample, the k closest to the sample is foundnearestOne sample, namely: if the sampleIn a sampleK of (a)nearestWithin a nearest neighbor, or sampleIn a sampleK of (a)nearestWithin a nearest neighbor, thenOtherwise, Wij＝0；Dii＝∑jWjiAnd L-D is a Laplacian matrix; sorting the calculated eigenvectors according to the magnitude of the eigenvalues, and taking the top k1A feature vectorLet Vi 1＝ai-1，i＝1,2,…,k1Then, thenNamely, the extracted convolution kernel is obtained;
a convolution kernel V1With each frame imageCarry out convolution, i.e. Then n is generated in the convolutional layertk1A web output feature map represented as:
(c) performing feature point downsampling on the obtained feature map Y based on maximum value pooling (maxporoling); let the sampling window size be s1×s1Then n is obtainedtk1Amplitude output feature map:
wherein, the ith output characteristic mapping chart Represents the jth row and kth column of the ith output characteristic map, u, v denote the sampling step size, YiRepresenting the ith input image, and max { } representing a maximum function; in addition, hereUsing non-overlapping samples, i.e. taking u-v-s1；
(d) Using a procedure similar to (b) for image XiPerforming block processing to set block size to p1×p2Taking the feature map Z obtained in the step (c) as the input of the convolutional layer, and removing the mean value of the block data of each frame of image to obtain an input image:wherein, the ith input feature mapRepresenting the ith block-de-averaged image,representing the jth block vector of the ith image after the mean value is removed; constructing a weight matrix W and according to ZLZTa＝λZDZTa, calculating eigenvectors, sorting according to the magnitude of eigenvalues, and taking the top k2The feature vector is used as the selected convolution kernelWherein, Vi 2，i＝1,…,k2Represents V2The ith convolution kernel of (a); the resulting convolution kernel V is then used2For each frame of imagePerforming convolution to generate n on the convolution layertk1k2Amplitude output feature map:wherein,
(e) for the feature map U obtained as described above, the same phase as (c) is usedA similar step, performing feature point down-sampling based on maximum value pooling; let the sampling window size be s2×s2Then n is obtainedtk1k2Amplitude output feature map:wherein, the ith output characteristic mapping chart Represents the jth row and kth column of the ith output characteristic map,u, v denote the sampling step size, UiRepresenting the ith input image, and max { } representing a maximum function; non-overlapping sampling is used here, i.e. taking u-v-s2；
(f) Order toI.e. taking each k in O2The images are grouped, and are subjected to Heaviside binary quantization and then are processed into decimal values, and each k is2Converting a picture into an image Wherein H (-) represents the Heaviside function, Pi jRepresents PiJ-th image of (1), TiRepresenting a decimal processing result, having a value range ofThen every k is taken1Breadth TiThe images are grouped, and each image is firstlyDividing the block into B blocks, calculating histogram features of each block region, and connecting the histogram features of the B blocks into row vectorsWherein,then for each image X in (a)lFinally extracting the feature vector based on the convolutional neural network
(3) Classification and identification: extracting the above-mentioned featuresTaking the target label corresponding to each feature vector as an input, and constructing by a multi-class Support Vector Machine (SVM) to obtain a classifier model of the target; based on the classifier model, the targets in different camera fields can be labeled and classified for target handover and tracking.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510047118.4A CN104616032B (en) | 2015-01-30 | 2015-01-30 | Multi-camera system target matching method based on depth convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510047118.4A CN104616032B (en) | 2015-01-30 | 2015-01-30 | Multi-camera system target matching method based on depth convolutional neural networks |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104616032A CN104616032A (en) | 2015-05-13 |
| CN104616032B true CN104616032B (en) | 2018-02-09 |
Family
ID=53150469
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510047118.4A Expired - Fee Related CN104616032B (en) | 2015-01-30 | 2015-01-30 | Multi-camera system target matching method based on depth convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104616032B (en) |
Families Citing this family (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104850836B (en) * | 2015-05-15 | 2018-04-10 | 浙江大学 | Insect automatic distinguishing method for image based on depth convolutional neural networks |
| CN107851174B (en) * | 2015-07-08 | 2021-06-01 | 北京市商汤科技开发有限公司 | Image semantic annotation equipment and method, and generation method and system of image semantic annotation model |
| CN105138973B (en) * | 2015-08-11 | 2018-11-09 | 北京天诚盛业科技有限公司 | The method and apparatus of face authentication |
| US10255040B2 (en) * | 2017-05-11 | 2019-04-09 | Veridium Ip Limited | System and method for biometric identification |
| CN105320961A (en) * | 2015-10-16 | 2016-02-10 | 重庆邮电大学 | Handwriting numeral recognition method based on convolutional neural network and support vector machine |
| CN105373796B (en) * | 2015-10-23 | 2019-01-25 | 河南眼神科技有限公司 | The method, apparatus and its application of image activation operation |
| CN106611177A (en) * | 2015-10-27 | 2017-05-03 | 北京航天长峰科技工业集团有限公司 | Big data-based image classification method |
| CN105354560A (en) * | 2015-11-25 | 2016-02-24 | 小米科技有限责任公司 | Fingerprint identification method and device |
| CN105719313B (en) * | 2016-01-18 | 2018-10-23 | 青岛邃智信息科技有限公司 | A kind of motion target tracking method based on intelligent real-time video cloud |
| WO2017206156A1 (en) * | 2016-06-03 | 2017-12-07 | Intel Corporation | Look-up convolutional layer in convolutional neural network |
| CN106203318B (en) * | 2016-06-29 | 2019-06-11 | 浙江工商大学 | Camera network pedestrian recognition method based on the fusion of multi-level depth characteristic |
| CN106227851B (en) * | 2016-07-29 | 2019-10-01 | 汤一平 | The image search method of depth of seam division search based on depth convolutional neural networks |
| CN106407891B (en) * | 2016-08-26 | 2019-06-28 | 东方网力科技股份有限公司 | Target matching method and device based on convolutional neural networks |
| US20180124437A1 (en) * | 2016-10-31 | 2018-05-03 | Twenty Billion Neurons GmbH | System and method for video data collection |
| CN106709441B (en) * | 2016-12-16 | 2019-01-29 | 北京工业大学 | A kind of face verification accelerated method based on convolution theorem |
| CN106504190B (en) * | 2016-12-29 | 2019-09-13 | 浙江工商大学 | A kind of three-dimensional video-frequency generation method based on 3D convolutional neural networks |
| CN110506277B (en) * | 2017-02-13 | 2023-08-08 | 诺基亚技术有限公司 | Filter reuse mechanism for constructing robust deep convolutional neural networks |
| CN106991428A (en) * | 2017-02-24 | 2017-07-28 | 中国科学院合肥物质科学研究院 | Insect image-recognizing method based on adaptive pool model |
| CN106980880A (en) * | 2017-03-06 | 2017-07-25 | 北京小米移动软件有限公司 | The method and device of images match |
| CN108572183B (en) | 2017-03-08 | 2021-11-30 | 清华大学 | Inspection apparatus and method of segmenting vehicle image |
| CN106991396B (en) * | 2017-04-01 | 2020-07-14 | 南京云创大数据科技股份有限公司 | Target relay tracking algorithm based on intelligent street lamp partner |
| CN107092935A (en) * | 2017-04-26 | 2017-08-25 | 国家电网公司 | A kind of assets alteration detection method |
| CN107393523B (en) * | 2017-07-28 | 2020-11-13 | 深圳市盛路物联通讯技术有限公司 | Noise monitoring method and system |
| CN107844795B (en) * | 2017-11-18 | 2018-09-04 | 中国人民解放军陆军工程大学 | Convolutional neural network feature extraction method based on principal component analysis |
| CN109146921B (en) * | 2018-07-02 | 2021-07-27 | 华中科技大学 | Pedestrian target tracking method based on deep learning |
| CN110320452A (en) * | 2019-06-21 | 2019-10-11 | 河南理工大学 | A kind of series fault arc detection method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104036323A (en) * | 2014-06-26 | 2014-09-10 | 叶茂 | Vehicle detection method based on convolutional neural network |
| CN104077613A (en) * | 2014-07-16 | 2014-10-01 | 电子科技大学 | Crowd density estimation method based on cascaded multilevel convolution neural network |
-
2015
- 2015-01-30 CN CN201510047118.4A patent/CN104616032B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104036323A (en) * | 2014-06-26 | 2014-09-10 | 叶茂 | Vehicle detection method based on convolutional neural network |
| CN104077613A (en) * | 2014-07-16 | 2014-10-01 | 电子科技大学 | Crowd density estimation method based on cascaded multilevel convolution neural network |
Non-Patent Citations (6)
| Title |
|---|
| Fully Convolutional Neural Networks for Crowd Segmentation;Kai Kang等;《Computer Science》;20141231;第49卷(第1期);第1-9页 * |
| Infrared moving target detection and tracking based on tensor locality preserving projection;Hong Li等;《Infrared Physics & Technology》;20090915;第77-83页 * |
| Palmprint Recognition via Locality Preserving Projections and Extreme Learning Machine Neural Network;Jiwen Lu等;《International Conference on Signal Processing》;20081208;第2096-2099页 * |
| Similar looking Gujarati printed character recognition using Locality Preserving Projection and Artificial Neural Networks;Mandar Chaudhary等;《2012 Third International Conference on Emerging Applications of Information Technology》;20121231;第153-156页 * |
| 基于局部保持投影和RBF神经网络的DOA估计;王荣秀等;《科学技术与工程》;20130831;第13卷(第24期);第7054-7058页 * |
| 基于误差投影和局部投影的RBF神经网络学习算法;宋绍云等;《玉溪师范学院学报》;20090430;第25卷(第4期);第46-50页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104616032A (en) | 2015-05-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104616032B (en) | Multi-camera system target matching method based on depth convolutional neural networks | |
| Huang et al. | Context-aware single image rain removal | |
| CN108256562B (en) | Salient target detection method and system based on weak supervision time-space cascade neural network | |
| Zhang et al. | Single-image crowd counting via multi-column convolutional neural network | |
| Zhang et al. | 4-dimensional local spatio-temporal features for human activity recognition | |
| CN106683119B (en) | Moving vehicle detection method based on aerial video image | |
| An et al. | Image super-resolution by extreme learning machine | |
| CN106127197B (en) | Image saliency target detection method and device based on saliency label sorting | |
| Yang et al. | ContourGAN: Image contour detection with generative adversarial network | |
| CN102542571B (en) | Moving target detecting method and device | |
| KR101436369B1 (en) | Apparatus and method for detecting multiple object using adaptive block partitioning | |
| CN110176024B (en) | Method, device, equipment and storage medium for detecting target in video | |
| Shukla et al. | Moving object tracking of vehicle detection: a concise review | |
| CN110929593A (en) | Real-time significance pedestrian detection method based on detail distinguishing and distinguishing | |
| Shijila et al. | Simultaneous denoising and moving object detection using low rank approximation | |
| CN110110618B (en) | SAR target detection method based on PCA and global contrast | |
| CN106157330A (en) | A kind of visual tracking method based on target associating display model | |
| CN115359407A (en) | Multi-vehicle tracking method in video | |
| Chen et al. | Attention-based hierarchical fusion of visible and infrared images | |
| Tsutsui et al. | Distantly supervised road segmentation | |
| CN109448024B (en) | Visual tracking method and system for constructing constraint correlation filter by using depth data | |
| CN109241932B (en) | Thermal infrared human body action identification method based on motion variance map phase characteristics | |
| CN109117850B (en) | Method for identifying corresponding infrared target image by utilizing visible light target image | |
| Bhinge et al. | Data-driven fusion of multi-camera video sequences: Application to abandoned object detection | |
| CN111160255B (en) | Fishing behavior identification method and system based on three-dimensional convolution network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB03 | Change of inventor or designer information | ||
| CB03 | Change of inventor or designer information |
Inventor after: Wang Huiyan Inventor after: Hua Jing Inventor before: Wang Huiyan Inventor before: Wang Xun Inventor before: He Xiaoshuang Inventor before: Chen Weigang |
|
| GR01 | Patent grant | ||
| 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: 20180209 Termination date: 20200130 |