CN108133192A - A kind of pedestrian based on Gauss-Laplace distribution statistics identifies again - Google Patents
A kind of pedestrian based on Gauss-Laplace distribution statistics identifies again Download PDFInfo
- Publication number
- CN108133192A CN108133192A CN201711430149.3A CN201711430149A CN108133192A CN 108133192 A CN108133192 A CN 108133192A CN 201711430149 A CN201711430149 A CN 201711430149A CN 108133192 A CN108133192 A CN 108133192A
- Authority
- CN
- China
- Prior art keywords
- sample
- difference
- distance
- represent
- gauss
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
- G06V40/23—Recognition of whole body movements, e.g. for sport training
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/30—Scenes; Scene-specific elements in albums, collections or shared content, e.g. social network photos or video
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Human Computer Interaction (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Health & Medical Sciences (AREA)
- Psychiatry (AREA)
- Social Psychology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of pedestrians based on Laplacian distribution statistics to identify again, feature is extracted from training sample by using craft or deep learning model first, the feature of sample pair is subtracted each other, positive sample is generated to difference and negative sample to difference, so as to obtain distributed constant of the positive sample to difference and negative sample to difference, then laplacian distribution is used to difference to positive sample, negative sample uses Gauss Distribution Fitting to difference, and then obtain distance metric function, finally inquiry picture and library picture are brought into distance metric function and calculate distance, it is ranked up according to the size of distance, obtain recognition result to the end.More accurately and reliably, expansion and applicability are stronger for distance metric function provided by the invention.
Description
Technical field
The invention belongs to pedestrian's weight identification technology fields, are related to a kind of pedestrian's weight identification technology side based on distance metric
Case more particularly to pedestrian's recognition methods again based on Gauss-Laplace distribution statistics.
Background technology
In pedestrian's weight identification technology field, pedestrian identifies that also referred to as pedestrian identifies again again, is to utilize computer vision technique
Judge with the presence or absence of the technology of specific pedestrian in image or video sequence, be exactly specifically to verify whether wrapped in two photos
The technology of personage is re-recognized containing same person or in a large amount of monitor camera.The technology can be in many keys
It is used in, such as multiple camera tracking, judicial evidence collection and video frequency search system.
Pedestrian's weight identification technology institute facing challenges mostly come from present, and query image and library image are seldom same
It is acquired in environment, so as to cause query image and library image, there are differences in resolution, illumination variation and viewpoint translation, these states
Variation to result in the work that pedestrian identifies again extremely difficult.Existing research is concentrated mainly on two aspects, and first is feature
The extraction of expression its object is to build the distinction visual descriptor for representing face or human body, i.e., is calculated from picture
Some feature representations with identification, obtain the digital expression about pedestrian's appearance.However, design one adapts to not co-occurrence
The powerful feature description of reality condition is challenging.Second is using some distance metric methods, is divided into two classes, first, iteration
Prioritization scheme, this kind of method from abundant exemplar learning priori by paying close attention to the design of loss function and rule
Change solution.However, almost all of algorithm is required for the optimized parameter of the metric function designed by iterative learning.Particularly
For large-scale dataset, the iterative calculation repeated is expensive.Therefore, when exemplar increases or needs online training,
This method is inappropriate.Second is that non-iterative statistical inference.This kind of method observes the generating process of sample pair.It was found that sample
To distribution after, give new sample pair, calculated according to the possibility of pairing difference.There is no iterative process, it is only necessary to
Distributed constant is calculated before authentication, therefore can expand to large data sets and online task.But traditional method exist away from
The problem of from measurement poor robustness, cause the accuracy rate that pedestrian identifies again low.
Chinese patent literature CN107330397A, open (bulletin) day 2017.11.07 disclose a kind of based between greatly
Every pedestrian's recognition methods again of relative distance metric learning, this method carries out dimension-reduction treatment to the feature representation vector of pedestrian image
And the vector after dimensionality reduction is further projected into subspace in class, then according to the feature representation vector after projection and corresponding mark
Label information learns mahalanobis distance metric matrix by optimizing loss function, and the metric matrix of acquisition has stronger robustness, carries
The accuracy rate that pedestrian identifies again is risen, but this method belongs to the distance metric method of iteration optimization class, Gauss-drawing is based on one kind
Recognition methods research angle is different to the pedestrian of this distribution statistics of pula again.
Chinese patent literature CN107316031A, open (day for announcing) 2017.11.03 are disclosed a kind of for pedestrian
The image characteristic extracting method identified again, the image characteristic extracting method described in this method refer to carry by being aligned local description
Global characteristics extraction is taken and be classified, compared with traditional feature extracting method, can be solved pedestrian identifies again in due to pedestrian's appearance
The feature that state variation etc. is brought is misaligned problem, eliminates the influence that the identification of extraneous background counterweight is brought, and thus improves pedestrian and knows again
Other precision and robustness, the non-iterative statistical inference technology not being related in distance metric are based on Gauss-drawing with one kind
Recognition methods research angle is different to the pedestrian of this distribution statistics of pula again.
Chinese patent literature CN107273873A, open (bulletin) day 2017.10.20 disclose one kind and are based on not advising
The then pedestrian of video sequence recognition methods and system again, this method is by detecting the stable point in condition curve, from video sequence
Multiple continuous subsequences are extracted, obtain candidate sequence, the reconstructed error of each subsequence is then asked for using rarefaction representation, is obtained
To each subsequence noise measurement as a result, then according to the noise measurement of each subsequence as a result, from candidate sequence cancelling noise
More than the subsequence of respective threshold, candidate pool is formed, finally carries out pedestrian's character representation of adaptive weighting, is obtained based on video
The retrieval result of sequence, improves the performance that pedestrian identifies again under irregular sequence, do not relate in distance metric it is non-repeatedly
For statistical inference technology, it is different to study angle from a kind of pedestrian based on Gauss-Laplace distribution statistics again recognition methods
's.
Invention content
In order to solve the shortcomings of the prior art, the present invention provides one kind to be based on Gauss-Laplace distribution statistics
Pedestrian's recognition methods again.
The technical solution adopted in the present invention is：A kind of pedestrian based on Gauss-Laplace distribution statistics side of identification again
Method includes the following steps that a kind of pedestrian based on Gauss-Laplace distribution statistics identifies again, which is characterized in that including following
Step：
Step 1：Feature is extracted from training sample；
Step 2：The feature of sample pair is subtracted each other, making positive sample, negative sample is to generating negative sample to generating positive sample to difference
This is to difference；
Step 3：Calculate distributed constant Σ of the positive sample to difference+With negative sample to the distributed constant Σ of difference-；
Step 4：Calculate distance metric function；
Step 5：The picture to prestore in picture and data library will be inquired to obtain inquiring picture and data library by step 1-2 principles
In prestore the difference of picture, difference is substituted into the distance metric function obtained by step 4, calculates and prestores in inquiry picture and data library
Picture distance, from small to large sort according to distance, distance it is minimum be same people maximum probability.
Compared with existing pedestrian weight identification technology, the present invention mainly has the following advantages and beneficial effect：
1) compared with prior art, the present invention proposes a kind of pedestrian based on Gauss-Laplace distribution statistics and knows again
Other learning distance metric method has sizable improvement compared with state-of-the-art iteration and non-iterative learn；
2) method compared with prior art, proposed is very simple, and operating cost is low, it means that non-iterative study away from
It verifies and applies more suitable for extensive line personnel from metric function；
Description of the drawings
Fig. 1 is the positive and negative differences between samples distribution map of the embodiment of the present invention；
Fig. 2 is the flow chart of the embodiment of the present invention；
Fig. 3 is the principle schematic of the embodiment of the present invention.
Specific embodiment
Understand for the ease of those of ordinary skill in the art and implement the present invention, it is right with reference to the accompanying drawings and embodiments
The present invention is described in further detail, it should be understood that and implementation example described herein is merely to illustrate and explain the present invention,
It is not intended to limit the present invention.
The present embodiment uses MATLAB R2015b and VS2013 as Simulation Experimental Platform, in order to assess proposed side
The applicability of method, the present embodiment are tested on 3 standard data sets with different various features.The present embodiment is first
Human digit is tested on widely used data set VIPeR, is then carried out on maximum data collection Market-1501
Assessment also has studied the facial problem of no constraint environment on LWF data sets.
VIPeR data sets include the outdoor images of 1,264 632 people, everyone has a pair of different from two respectively
The image of camera shooting, all images are normalized to the size of 128 × 48 pixels, visual angle change, illumination and picture quality
Variation is very big.The present embodiment lists positive sample pair of some two pictures from common identity and from different identity
Some negative samples pair.The present embodiment follows general assessment agreement, by 632 images to being randomly divided into two groups, every group of 316 images
Right, one is used to train, another calculates the CMC (Cumulative for the different stage for running more than 5 times for testing
Matching Characteristic) value.CMC value refers in n times inquiry, there is correct pedestrian's object before return in r result
Probability, when r result before return, CMC value is higher, represents that pedestrian retrieval performance is better.
Market-1501 data sets include the bounding box of 1501 identity of 32668 marks.The image of each identity is most
Mostly shot by six video cameras.According to known to the present embodiment, it is that people maximum so far re-recognizes data set.The present embodiment
The general assessment agreement of this data set is followed, calculates the CMC value of different stage, while calculate mAP (mean Average
Precision) value.
LFW data sets include the 13000 of 5749 subjects several face-images collected from network, expression, posture, year
The variation of age, illumination, resolution ratio etc. is very big.The present embodiment test the present embodiment under " image is limited " setting of standard
Algorithm, the setting is dedicated for verification.In this setting, data set is divided into 10 completely self-contained files, it is ensured that different
People occur on different files.300 erect images and 300 negative-appearing images pair are provided in each file.Each the present embodiment is 9
Training obtains the distance metric function of the present embodiment in a file, then assesses the picture in remaining file.It calculates after this
The Mean match precision of 10 times is final result.
See Fig. 2 and Fig. 3, a kind of pedestrian based on Gauss-Laplace distribution statistics provided by the invention identifies again, packet
Include following steps：
Step 1：Feature is extracted from training sample by using craft or deep learning model；
The prior art can be used, such as on VIPeR data sets in Feature Extraction Technology specific implementation pedestrian identifies again in
Feature SCNCD, LOMO of the manual extraction of use and based on deep learning model FTCNN, are adopted on Market-1501 data sets
With based on CaffeNet and ResNet-50, used on LFW data sets and feature extraction is carried out based on Lightened CNN.Specifically
During implementation, document can refer to：
Yang Yang,Jimei Yang,Junjie Yan,Shengcai Liao,Dong Yi,and Stan
Z.Li.2014.Salient Color Names for Person Reidentification.In ECCV.
Shengcai Liao,Yang Hu,Xiangyu Zhu,and Stan Z.Li.2015.Person re-
identification by Local Maximal Occurrence representation and metric
learning.In CVPR.
Tetsu Matsukawa and Einoshin Suzuki.2016.Person reidentification
using cnn features learned from combination of attributes.In ICPR.
Liang Zheng,Yi Yang,and Alexander G.Hauptmann.2016.Person Re-
identification:Past,Present and Future.In arXiv.
Xiang Wu,Ran He,Zhenan Sun,and Tieniu Tan.2015.A Light CNN for Deep
Face Representation with Noisy Labels.In arXiv.
Those skilled in the art can voluntarily specific pedestrian's weight identification feature extractive technique selected to use, the present invention be not superfluous
It states.
Step 2：The feature of sample pair is subtracted each other, making positive sample, negative sample is to generating negative sample to generating positive sample to difference
This is to difference；
Step 3：Calculate distributed constant Σ of the positive sample to difference+With negative sample to the distributed constant Σ of difference-；
Calculate distributed constant Σ+And Σ-, carry out side's calculating respectively to difference to difference and negative sample to positive sample, obtain
The distributed constant of positive sample pair and negative sample pair.Parameter Σ+With Σ-realization method it is as follows,
If by the training sample obtained by step 1 to for (xi,xj), then Σ+With Σ-be defined as follows,
Wherein, yijRepresent training sample to (xi,xj) classification, training sample is to (xi,xj) be positive sample pair when yij=1,
Y during for negative sample pairij=0, N+Represent the sum of positive sample pair, N-Represent the sum of negative sample pair；
Step 4：Calculate distance metric function；
Distance metric function is calculated, the probability when distance of sample pair is set as the sample to being negative sample pair and the sample
The ratio of probability during to being positive sample pair.Analysis finds that positive sample follows difference laplacian distribution, negative sample after tested
Gaussian Profile is followed to difference.
This is the discovery that is obtained by Preliminary visualization experiment and the test of fitness of fot.The data of visualized experiment are from LFW
It is obtained in data set, comprising 3000 positive samples pair and 3000 negative samples pair, with trained Lightened CNN model extractions
Go out 256 dimensional features of each image, each sample is to be represented by the difference of individual features to difference.Therefore, the present embodiment obtains
The positive sample different to 3000 is to difference and 3000 different negative samples to difference.
The present embodiment visualizes the two distributions in two-dimensional space, in order to which the difference distribution to different dimensions is compared
Comprehensive observation, the present embodiment select key dimension by random and principal component analysis.First, the present embodiment random selection two
A dimension checks the overall distribution figure of difference, as shown in Fig. 1 (a) and Fig. 1 (b).Fig. 1 (a) and Fig. 1 (b) are respectively illustrated just
Sample is to difference and negative sample to the distribution map of difference.Distribution of the positive sample to difference is can be seen that from the shape of this two width figure
It is centrally located place's spike difference higher compared with the negative sample difference distribution figure for following Gaussian Profile, it appears that unlike former
The Gaussian Profile of hypothesis and be more likely to laplacian distribution.Secondly, the selection that the present embodiment passes through principal component analysis (PCA)
Two key dimensions check the distribution map of difference.Fig. 1 (c) and Fig. 1 (d) respectively illustrate positive sample under PCA to difference and
Negative sample is to the distribution map of difference, it can be found that showing the phenomenon that identical with Fig. 1 (a) and Fig. 1 (b).
As can be seen from Figure 1, positive sample seems to be more likely to unlike the Gaussian Profile assumed in the past to the distribution of difference
Laplacian distribution.In order to verify this observation, the present embodiment is further fitted different distributions using Minitab tools
Goodness is examined, and the respective dimensions of all differences is selected to study their distribution with Minitab, obtain the result of table 1-4.Table 1-
4 respectively illustrate the negative sample of a dimension and all dimensions to the test of fitness of fot of difference and positive sample to difference distribution
Value.AD (Anderson-Darling) counts whether amount measurement data follows specific distribution in table, is distributed more suitable data, AD
Statistic is with regard to smaller.In addition, if the P values (if available) of AD tests are horizontal higher than selected significant property (being usually 0.05),
Then it could be assumed that the data follow the distribution specified.The P of Gaussian Profile in table 1 and table 3>0.05 and AD values are minimum, this shows
Negative sample meets Gaussian Profile very much to the distribution of difference.The AD values of exponential distribution are the minimum values in table in table 2 and table 4, this
Mean that positive sample is to the half of the distribution of difference it is more likely that an index compared with other are distributed.Due to Laplce point
Cloth is considered two exponential distributions being stitched together, so the present embodiment thinks overall distribution of the positive sample to difference
It is more likely that laplacian distribution.
The test of fitness of fot that 1 negative sample of table ties up difference 1
| Distribution | AD | P |
| Gaussian | 0.658 | 0.086 |
| Exponential | 714.426 | <0.01 |
| Smallest Extreme Value | 51.418 | <0.01 |
| Largest Extreme Value | 38.363 | <0.01 |
| Logistic | 1.348 | <0.005 |
The test of fitness of fot that 2 positive sample of table ties up difference 1
| Distribution | AD | P |
| Gaussian | 48.657 | <0.05 |
| Exponential | 8.658 | <0.05 |
| Smallest Extreme Value | 94.172 | <0.01 |
| Largest Extreme Value | 10.241 | <0.01 |
| Logistic | 25.178 | <0.005 |
3 negative sample of table examines average fit goodness of the difference in all dimensions
| Distribution | AD | P |
| Gaussian | 0.612 | 0.092 |
| Exponential | 701.124 | <0.01 |
| Smallest Extreme Value | 3.312 | <0.01 |
| Largest Extreme Value | 40.112 | <0.01 |
| Logistic | 1.945 | <0.005 |
4 positive sample of table examines average fit goodness of the difference in all dimensions
| Distribution | AD | P |
| Gaussian | 49.287 | <0.05 |
| Exponential | 7.158 | <0.05 |
| Smallest Extreme Value | 98.123 | <0.01 |
| Largest Extreme Value | 13.932 | <0.01 |
| Logistic | 26.291 | <0.005 |
Distance metric model is set based on this discovery, measures characteristic is by parameter Σ+And Σ-It determines.According to positive sample
Meet difference laplacian distribution, negative sample meets Gaussian Profile to difference, the sample pair based on Gauss-Laplace distribution
Range formula δ (xi,xj) be defined as follows：
Wherein, fG(xij| Σ -) it represents to assume sample to (xi,xj) be negative sample pair probability density function, G represents negative sample
This follows Gaussian Profile to difference；Similarly, fL(xij|Σ+) represent to assume sample to (xi,xj) be positive sample pair probability density
Function, L represent that negative sample follows difference laplacian distribution, and two probability density functions represent as follows：
Wherein Km(x) it represents the modified Bessel function of the second class, is calculated at x；D represents the dimension of sample characteristics；λ ＞
0 represents scale parameter；Function ψ (xij,Σ+) and ψ (xij,Σ-) be defined as follows：
By fG(xij|Σ-) and fL(xij|Σ+) substitute into δ (xi,xj) in, then take the logarithm, obtain sample and adjust the distance δ (xi,xj) full
The following distribution of foot：
Above formula is reduced to：
The equation substantially defines the distance metric function based on Gauss-Laplace distribution statistics, measures characteristic
By parameter Σ+And Σ-It determines.
Step 5：The picture to prestore in picture and data library will be inquired to obtain in inquiry picture and data library in advance by step 1-2
The difference of picture is deposited, difference is substituted into the distance metric function obtained by step 4, calculates the figure to prestore in inquiry picture and data library
The distance of piece from small to large sorts according to distance, distance it is minimum be same people maximum probability.
In the above process, subtracted each other by feature extraction and feature, generate positive sample to difference and negative sample to difference, thus
Obtain distributed constant of the positive sample to difference and negative sample to difference, then to positive sample to difference using laplacian distribution,
Negative sample uses Gauss Distribution Fitting to difference, and then obtains distance metric function, will finally inquire picture and library picture is brought into
Distance is calculated in distance metric function, is ranked up according to the size of distance, according to the different evaluation criteria of each data set, is obtained
It is last as a result, being shown in Table 5-7.It can be found that the pedestrian based on Gauss-Laplace distribution statistics of the present invention knows again from table
Other method (GL) retrieval accuracy higher.
CMC of the table 5 on LFW data sets
| Method | 1 | 5 | 10 | 20 |
| PRDC (document 1) | 15.7 | 38.4 | 53.9 | 70.1 |
| SDALF (document 2) | 19.9 | 38.4 | 49.4 | 66 |
| BiCov (document 3) | 20.6 | 43.2 | 56.1 | 68 |
| ESDC (document 4) | 26.3 | 46.4 | 58.6 | 72.8 |
| DeepMetric (document 5) | 28.2 | 59.3 | 73.4 | 86.4 |
| LADF (document 6) | 30 | 64 | 80 | 92 |
| CPDL (document 7) | 34 | 64.2 | 77.5 | 88.6 |
| MKML (document 8) | 37 | 69.9 | 80.7 | 90.1 |
| DeepFeature (document 9) | 40.5 | 60.8 | 70.4 | 84.4 |
| SCNCD+KISSME (document 10) | 34.81 | 66.77 | 80.38 | 89.56 |
| SCNCD+GL (the method for the present invention) | 38.92 | 69.3 | 81.33 | 90.51 |
| LOMO+XQDA (document 11) | 40 | 68.13 | 80.51 | 91.08 |
| LOMO+KISSME | 28.48 | 57.59 | 76.58 | 89.24 |
| LOMO+GL (the method for the present invention) | 35.76 | 65.19 | 78.16 | 91.14 |
| FTCNN+XQDA | 31.2 | 59.8 | 74 | 83.5 |
| FTCNN+KISSME | 31.53 | 62.86 | 78.37 | 88.18 |
| FTCNN+GL (the method for the present invention) | 40.39 | 67.97 | 81.9 | 91.66 |
CMC and mAP of the table 6 on LFW data sets
| Method | 1 | 10 | 50 | mAP |
| BiCov (document 3) | 8.28 | - | - | 2.23 |
| LOMO | 26.07 | - | - | 7.75 |
| BoW (document 12) | 35.84 | 60.33 | 75.8 | 14.75 |
| Siamese CNN (document 13) | 65.88 | - | - | 39.5 |
| LSTM (document 14) | 61.6 | - | - | 35.3 |
| SCSP (document 15) | 51.9 | - | - | 26.35 |
| LOMO+NFST (document 16) | 55.43 | - | - | 29.87 |
| CaffeNet+L2 | 59.53 | 85.51 | 94.9 | 32.85 |
| CaffeNet+XQDA | 61.34 | 87.26 | 95.13 | 37.59 |
| CaffeNet+KISSME | 61.46 | 86.55 | 94.8 | 36.63 |
| CaffeNet+GL (the method for the present invention) | 62.08 | 86.82 | 95.01 | 36.86 |
| ResNet+L2 | 75.62 | 91.89 | 96.97 | 50.68 |
| ResNet+XQDA | 75.5 | 91.66 | 96.94 | 52.87 |
| ResNet+KISSME | 77.52 | 92.93 | 97.48 | 53.87 |
| ResNet+GL (the method for the present invention) | 78.03 | 93.44 | 97.68 | 53.76 |
Matching precision of the table 7 on LFW data sets
Note：Bibliography is：
[document 1] Wei Shi Zheng, Shaogang Gong, and Tao Xiang.2011.Person re-
identification by probabilistic relative distance comparison.In CVPR.
[document 2] M.Farenzena, L.Bazzani, A.Perina, V.Murino, and
M.Cristani.2010.Person re-identification by symmetry-driven accumulation of
local features.In CVPR.
[document 3] Bingpeng Ma, Yu Su, and Frdric Jurie.2014.Covariance descriptor
based on bio-inspired features for person re-identification and face
verification.Image Vision Comput.(2014).
[document 4] Rui Zhao, Wanli Ouyang, and Xiaogang Wang.2013.Unsupervised
Salience Learning for Person Re-identification.In CVPR.
[document 5] Dong Yi, Zhen Lei, Shengcai Liao, and S.Z Li.2014.Deep Metric
Learning for Person Re-identification.In ICPR.
[document 6] Zhen Li, Shiyu Chang, Feng Liang, Thomas S Huang, Liangliang Cao,
and John R Smith.2013.Learning Locally-Adaptive Decision Functions for Person
Verification.In CVPR.
[document 7] Sheng Li, Ming Shao, and Yun Fu.2015.Cross-View Projective
Dictionary Learning for Person Re-identification.In IJCAI.
[document 8] Muhammad Adnan Syed and Jianbin Jiao.2016.Multi-kernel metric
learning for person re-identification.In ICIP.
[document 9] Shengyong Ding, Liang Lin, Guangrun Wang, and Hongyang
Chao.2015.Deep feature learning with relative distance comparison for person
re-identification.Pattern Recogn.(2015).
[document 10] Martin K ¨ ostinger, Martin Hirzer, Paul Wohlhart, and Peter
M.Roth.2012.Large scale metric learning from equivalence constraints.In CVPR.
[document 11] Shengcai Liao, Yang Hu, Xiangyu Zhu, and Stan Z.Li.2015.Person
re-identification by Local Maximal Occurrence representation and metric
learning.In CVPR.
[document 12] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang,
and Qi Tian.2015.Scalable Person Re-identification:A Benchmark.In ICCV.
[document 13] Rahul Rama Varior, Mrinal Haloi, and Gang Wang.2016.Gated
Siamese Convolutional Neural Network Architecture for Human Re-
Identification.In ECCV.
[document 14] Rahul Rama Varior, Bing Shuai, Jiwen Lu, Dong Xu, and Gang
Wang.2016.A Siamese Long Short-Term Memory Architecture for Human Re-
identification.In ECCV.
[document 15] Dapeng Chen, Zejian Yuan, Badong Chen, and Nanning
Zheng.2016.Similarity Learning with Spatial Constraints for Person Re-
identification.In CVPR.
[document 16] Li Zhang, Tao Xiang, and Shaogang Gong.2016.Learning a
Discriminative Null Space for Person Re-identification.In CVPR.
[document 17] Gary B.Huang, Michael J.Jones, and Eric Learned-Miller.2008.LFW
results using a combined Nowak plus MERL recognizer.In ECCV workshop.
[document 18] Hieu V.Nguyen and Li Bai.2010.Cosine Similarity Metric
Learning for Face Verification.In ACCV.
[document 19] Yiming Ying and Peng Li.2012.Distance Metric Learning with
Eigenvalue Optimization.J.Mach.Learn.Res.(2012).
[document 20] G.B.Huang, Honglak Lee, and E Learned-Miller.2012.Learning
hierarchical representations for face verification with convolutional deep
belief networks.In CVPR.
[document 21] Junlin Hu, Jiwen Lu, and Yap Peng Tan.2014.Discriminative Deep
Metric Learning for Face Verification in the Wild.In CVPR.
[document 22] Xiangyu Zhu, Zhen Lei, Junjie Yan, and Dong Yi.2015.Highfidelity
Pose and Expression Normalization for face recognition in the wild.In CVPR.
It should be understood that the part that this specification does not elaborate belongs to the prior art.
It should be understood that the above-mentioned description for preferred embodiment is more detailed, can not therefore be considered to this
The limitation of invention patent protection range, those of ordinary skill in the art are not departing from power of the present invention under the enlightenment of the present invention
Profit is required under protected ambit, can also be made replacement or deformation, be each fallen within protection scope of the present invention, this hair
It is bright range is claimed to be determined by the appended claims.
Claims (4)
1. a kind of pedestrian based on Gauss-Laplace distribution statistics identifies again, which is characterized in that includes the following steps：
Step 1：Feature is extracted from training sample；
Step 2：The feature of sample pair is subtracted each other, making positive sample, negative sample is to generating negative sample pair to generating positive sample to difference
Difference；
Step 3：Calculate distributed constant Σ of the positive sample to difference+With negative sample to the distributed constant Σ of difference-；
Step 4：Calculate distance metric function；
Step 5：The picture to prestore in picture and data library will be inquired to obtain in inquiry picture and data library in advance by step 1-2 principles
The difference of picture is deposited, difference is substituted into the distance metric function obtained by step 4, calculates the figure to prestore in inquiry picture and data library
The distance of piece from small to large sorts according to distance, distance it is minimum be same people maximum probability.
2. the pedestrian according to claim 1 based on Gauss-Laplace distribution statistics identifies again, it is characterised in that：Step
In rapid 1, feature is extracted from training sample by using craft or deep learning model.
3. the pedestrian according to claim 1 based on Gauss-Laplace distribution statistics identifies again, it is characterised in that：Step
In rapid 3,
Wherein, (xi,xj) represent training sample pair, yijRepresent training sample to (xi,xj) classification, training sample is to (xi,xj)
Y during for positive sample pairij=1, y when being negative sample pairij=0, N+Represent the sum of positive sample pair, N-Represent the total of negative sample pair
Number.
4. the pedestrian according to claim 3 based on Gauss-Laplace distribution statistics identifies again, it is characterised in that：Step
In rapid 4, to positive sample to difference using laplacian distribution, negative sample to difference using Gauss Distribution Fitting, and then obtain away from
From metric function δ (xi,xj)；
Wherein, fG(xij|Σ-) represent training sample to (xi,xj) be negative sample pair probability density function, G represent negative sample pair
Difference follows Gaussian Profile；fL(xij|Σ+) represent training sample to (xi,xj) be positive sample pair probability density function, L represent
Negative sample follows laplacian distribution to difference；Km(x) it represents the modified Bessel function of the second class, is calculated at x；D is represented
The dimension of sample characteristics；λ ＞ 0 represent scale parameter；
(5), (6), (7), (8) formula are substituted into distance metric function δ (xi,xj) in, it takes the logarithm, obtains sample and adjust the distance δ (xi,xj) full
The following distribution of foot：
Sample is adjusted the distance δ (xi,xj) final mathematical expression form is：
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711430149.3A CN108133192A (en) | 2017-12-26 | 2017-12-26 | A kind of pedestrian based on Gauss-Laplace distribution statistics identifies again |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711430149.3A CN108133192A (en) | 2017-12-26 | 2017-12-26 | A kind of pedestrian based on Gauss-Laplace distribution statistics identifies again |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108133192A true CN108133192A (en) | 2018-06-08 |
Family
ID=62393063
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711430149.3A Pending CN108133192A (en) | 2017-12-26 | 2017-12-26 | A kind of pedestrian based on Gauss-Laplace distribution statistics identifies again |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108133192A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108549883A (en) * | 2018-08-06 | 2018-09-18 | 国网浙江省电力有限公司 | A kind of face recognition methods again |
| CN109376578A (en) * | 2018-08-27 | 2019-02-22 | 杭州电子科技大学 | A kind of small sample target identification method based on depth migration metric learning |
| CN110428023A (en) * | 2019-05-31 | 2019-11-08 | 武汉大学 | A kind of counterreconnaissance escape attack method towards depth pedestrian weight identifying system |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107145826A (en) * | 2017-04-01 | 2017-09-08 | 浙江大学 | The pedestrian's recognition methods again reordered based on double constraint metric learnings and sample |
-
2017
- 2017-12-26 CN CN201711430149.3A patent/CN108133192A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107145826A (en) * | 2017-04-01 | 2017-09-08 | 浙江大学 | The pedestrian's recognition methods again reordered based on double constraint metric learnings and sample |
Non-Patent Citations (2)
| Title |
|---|
| ZHENG WANG ETAL.: "Statistical Inference of Gaussian-Laplace Distribution for Person Verification", 《PROCEEDINGS OF THE 25TH ACM INTERNATIONAL CONFERENCE ON MULTIMEDIA》 * |
| 苏松志等: "《行人检测：理论与实践》", 31 March 2016 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108549883A (en) * | 2018-08-06 | 2018-09-18 | 国网浙江省电力有限公司 | A kind of face recognition methods again |
| CN109376578A (en) * | 2018-08-27 | 2019-02-22 | 杭州电子科技大学 | A kind of small sample target identification method based on depth migration metric learning |
| CN110428023A (en) * | 2019-05-31 | 2019-11-08 | 武汉大学 | A kind of counterreconnaissance escape attack method towards depth pedestrian weight identifying system |
| CN110428023B (en) * | 2019-05-31 | 2021-09-14 | 武汉大学 | Anti-reconnaissance escape attack method for deep pedestrian re-identification system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111709311B (en) | Pedestrian re-identification method based on multi-scale convolution feature fusion | |
| Jourabloo et al. | Large-pose face alignment via CNN-based dense 3D model fitting | |
| CN101510257B (en) | Human face similarity degree matching method and device | |
| CN105224937B (en) | Fine granularity semanteme color pedestrian recognition methods again based on human part position constraint | |
| CN102906787A (en) | Facial analysis techniques | |
| CN101131728A (en) | Face shape matching method based on Shape Context | |
| Molina-Moreno et al. | Efficient scale-adaptive license plate detection system | |
| CN112016605A (en) | Target detection method based on corner alignment and boundary matching of bounding box | |
| CN104392246A (en) | Inter-class inner-class face change dictionary based single-sample face identification method | |
| Fang et al. | Real-time hand posture recognition using hand geometric features and fisher vector | |
| CN110852152A (en) | Deep hash pedestrian re-identification method based on data enhancement | |
| CN108133192A (en) | A kind of pedestrian based on Gauss-Laplace distribution statistics identifies again | |
| CN104376312B (en) | Face identification method based on bag of words compressed sensing feature extraction | |
| Jia et al. | Semantic and context based image retrieval method using a single image sensor for visual indoor positioning | |
| CN110852292A (en) | Sketch face recognition method based on cross-modal multi-task depth measurement learning | |
| CN104463085A (en) | Face recognition method based on local binary pattern and KFDA | |
| CN111144469B (en) | End-to-end multi-sequence text recognition method based on multi-dimensional associated time sequence classification neural network | |
| CN104200222B (en) | Object identifying method in a kind of picture based on factor graph model | |
| Zhang et al. | A multi-view camera-based anti-fraud system and its applications | |
| Campos et al. | Global localization with non-quantized local image features | |
| Tang et al. | A GMS-guided approach for 2D feature correspondence selection | |
| CN115050044A (en) | Cross-modal pedestrian re-identification method based on MLP-Mixer | |
| Yong et al. | Video-based detection of abnormal behavior in the examination room | |
| Sala et al. | 3-d volumetric shape abstraction from a single 2-d image | |
| Xudong et al. | Pedestrian detection and tracking with deep mutual learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180608 |