CN109271852A - A kind of processing method that the pedestrian detection based on deep neural network identifies again - Google Patents
A kind of processing method that the pedestrian detection based on deep neural network identifies again Download PDFInfo
- Publication number
- CN109271852A CN109271852A CN201810888879.6A CN201810888879A CN109271852A CN 109271852 A CN109271852 A CN 109271852A CN 201810888879 A CN201810888879 A CN 201810888879A CN 109271852 A CN109271852 A CN 109271852A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- network
- sample
- rcnn
- faster
- 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
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Human Computer Interaction (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of processing methods that the pedestrian detection based on deep neural network identifies again, it is on the basis of Faster-RCNN object detection network, according to the following steps improve: step 1, improve Faster-RCNN:1), Faster-RCNN is added to an additional region recommendation network, 2) it, is finally added to the full articulamentum of 256 neurons in improved network for extracting the relevant feature of pedestrian's identity, and adds calculating of the characteristic storage module for loss function;3) On-line matching loss function OLP and the preferential loss function HEP of difficult sample, are added to after the full articulamentum identified again for pedestrian;The training of Faster-RCNN after step 2, improvement;The test of Faster-RCNN after step 3, improvement.The solution have the advantages that: it is integrated together first is that pedestrian detection is identified again with pedestrian, improves the performance that the pedestrian based on pedestrian detection network identifies again;Second is that improving the accuracy of pedestrian's search mission.
Description
Technical field
The invention belongs to pedestrian detections and pedestrian to identify field again.
Background technique
With sharply increasing for camera head monitor data, pedestrian detection with pedestrian come into being again by identification technology.Pedestrian's inspection
Survey technology is mainly used in intelligent driving, auxiliary drives and the related fieldss such as intelligent monitoring, and pedestrian again widely answer by identification technology
It is monitored for criminal investigation, field of image search." pedestrian detection " main purpose is to whether there is pedestrian in detection image or video,
Without judging the pedestrian and whether other pedestrians belong to the same pedestrian, and " pedestrian identifies again " is also known as that " pedestrian searches
Rope ", main purpose are to judge whether some pedestrian in some camera once appeared in other cameras, that is, need by
Some pedestrian's feature is compared with other pedestrian's features, judges whether to belong to the same pedestrian.Solving pedestrian's search mission
When, existing method identifies pedestrian detection and pedestrian again as two separation the step of progress, and identification method is all again by pedestrian at present
It is based on the pedestrian image extracted.
In the pedestrian detection of actual monitored, face's effective information of pedestrian can not be captured, usually using the whole of pedestrian
Body information is identified again.And in pedestrian again identification process, due to the posture of pedestrian, illumination, camera angle etc. is multiple
The influence of factor may make the aspect ratio of different pedestrians increasingly similar with the feature of a group traveling together, cause pedestrian in this way and identify again
The problems such as there are erroneous detections.
Learning better feature representation is a kind of relatively effective mode, and the concept of deep learning is derived from artificial neural network
Research.Multilayer perceptron containing more hidden layers is exactly a kind of deep learning structure.Deep learning is formed by combining low-level feature
More abstract high-level characteristic indicates, to solve complicated computer vision problem.
Depth convolutional network is exactly the machine learning model under a kind of supervised learning, and the basic step of training and test is such as
Under:
1, prepare data, prepare training and test data with corresponding label;
2, ready training data is sent into network to be trained, utilizes stochastic gradient descent (SGD) when training
Network parameter is optimized.According to Bouvrie, J..Notes on convolutional neural networks. is (deep
Spend the explanation of convolutional network) BP algorithm recorded in Neural Nets. can parameter to each layer in deep neural network into
Row derivation (calculates).Assuming that the loss function of network are as follows:WhereinFor loss function, f () is
The function of the fitting of neural network, xi, w is respectively the parameter of input sample and neural network, yiFor the label of sample.Each
Sample seeks partial derivative to w to update the parameter of network;
3, after training to network convergence, network is inputted using test set sample, calculates the output result of network simultaneously
And be compared with true tag, the result of network may finally be tested out.
" Faster-RCNN:Towards Real-Time Object Detection with Region Proposal
Networks " (" Faster-RCNN: carrying out real-time object detection using region recommendation network ") Shaoqing Ren, Kaiming
He, Ross Girshick, and Jian Sun, International Conference on Neural Information
Processing Systems.MIT Press, 2015:91-99 describe Faster-RCNN object detection network, it is one
Object detection network structure based on deep learning, after inputting a picture, Faster-RCNN object detection network can be with
Export object category belonging to detection block coordinate and detection block.Firstly, picture of the network according to input, RPN sub-network can generate
It is a large amount of that frame proposal is recommended to be used for subsequent detection and classification task, then, recommend the ROIpooling meeting of pool area layer
Extract relevant feature and (whether being examined object) identified and classified to detection object, and to the detection block of object into
Row amendment.
The present invention is improved for Faster-RCNN object detection network, to realize that pedestrian detection identifies again with pedestrian,
Improve the accuracy that pedestrian identifies again.
" pedestrian detection identifies again " described in present patent application refers to the integration that pedestrian detection and pedestrian identify again, input
It include the pictures of same target pedestrian, the position coordinates for exporting pedestrian in picture for two, and to each row detected
People's output is one 256 dimension pedestrian identification feature again;The function of pedestrian detection is scanned for target pedestrian, and output is suggested mentioning
Frame is taken, the function that pedestrian identifies again is to carry out feature extraction to the proposed extraction frame of pedestrian detection output and compare.
Summary of the invention
The technical problem to be solved by the invention is to provide a kind of pedestrian detections based on deep neural network to identify again
Processing method, pedestrian detection is identified again with pedestrian and is integrated together by it, not only convenient for executing pedestrian's identification mission again, but also is promoted
The performance that pedestrian based on pedestrian detection network identifies again improves the accuracy of pedestrian's search mission.
Insight of the invention is that constructing a kind of pedestrian detection end to end and pedestrian identifies the network structure of combination, institute again
Call it is end-to-end, be exactly by pedestrian detection network and pedestrian again identification mission by depth Network integration to together, directly from picture
Target person is found in scene without artificial cutting image, according to the detection part of network in Faster-RCNN model, life
At window is suggested, it is put into subsequent network and carries out feature extraction and metric learning is carried out to feature by loss function.
In order to solve the above technical problems, the present invention is on the basis of Faster-RCNN object detection network, utilization is following
Step improves network structure:
Step 1 improves Faster-RCNN
1) Faster-RCNN, is added to an additional region recommendation network RPN, makes improved Faster-RCNN net
Network can input two pictures simultaneously, and obtain the corresponding recommendation region of each picture;
2), in network finally, the full articulamentum for being added to 256 neurons is relevant for extracting pedestrian's identity
Feature, and add calculating of the characteristic storage module for loss function;
3) it, is added to On-line matching loss function OLP after the full articulamentum identified again for pedestrian and difficult sample is excellent
First loss function HEP learns for knowing another characteristic again to pedestrian;
The training of Faster-RCNN after step 2, improvement
Two picture inputs containing identical pedestrian are improved Faster-RCNN network, recommend net using the region of two-way
Network RPN extracts the recommendation frame proposals of two road networks respectively, recycles and recommends pool area layer ROIpooling convolutional layer
The upper feature for recommending frame corresponding position is sent into full articulamentum, and full articulamentum is to recommending frame to carry out further screening, amendment, simultaneously
Pedestrian's identification feature again is extracted, loss function HEP addition On-line matching loss function OLP preferential with difficult sample is for extracting
Pedestrian behind the 256 full articulamentums of dimension of identification feature, supervises the study of whole network again;
The test of Faster-RCNN after step 3, improvement
A test picture is inputted into improved Faster-RCNN, is calculated, is obtained using trained network parameter
Final pedestrian detection result and pedestrian identify required feature again out.
Compared with the existing methods, the invention has the following advantages that
1, pedestrian detection and pedestrian are identified integration by the present invention again, provide a kind of new solution for pedestrian's search;
2, the present invention improves the accuracy rate of pedestrian's search.
Detailed description of the invention
Detailed description of the invention of the invention is as follows:
Fig. 1 is the structure simplification figure for improving Faster-RCNN;
Fig. 2 is the schematic diagram of OLP loss function.
Specific embodiment
Present invention will be further explained below with reference to the attached drawings and examples:
The present invention is improved according to the following steps on the basis of Faster-RCNN object detection network:
Step 1 improves Faster-RCNN
It is used as input as shown in Figure 1, improving Faster-RCNN and receiving two pictures comprising common identity pedestrian simultaneously,
The region recommendation network RPN that two pictures respectively enter the shared weighting parameter of two-way is calculated, while obtaining two pictures
Recommend frame, then the corresponding feature of recommendation frame region of two pictures is respectively fed to pool area layer ROIpooling is recommended to carry out
Pondization operation, the full articulamentum treated feature being sent into after ROIpooling is further processed, final to utilize
The full articulamentum of one 256 dimension is extracted knows another characteristic for pedestrian again, while exporting the coordinate of detection block and corresponding to detection block
Score.
Improving Faster-RCNN is to have 3 points of improvement on the basis of Faster-RCNN object detection network:
1) RPN of single channel, is improved to the RPN that two-way shares weight;
2), after recommending pool area layer, it is added to the full articulamentum of 256 new neurons, for extracting pedestrian
The correlated characteristic identified again;
3), after for extracting the full articulamentum that pedestrian identifies again, setting On-line matching loss function OLP and difficult sample
Preferential loss function HEP, for supervising the training of pedestrian's identification feature again.
The training of Faster-RCNN after step 2, improvement
The picture with identical pedestrian is matched according to the detection block of data set mark and corresponding label first, is being instructed
In experienced each iterative process, the picture matched is inputted in pairs in improved Faster-RCNN;
Two pictures after the recommendation network of region, obtain the coordinate of the recommendation frame of two pictures respectively, utilize respectively
Recommend pool area layer convolutional layer feature corresponding to recommendation frame, be input to full articulamentum, is used for using the extraction of full articulamentum
Pedestrian knows another characteristic again;
Loss function HEP On-line matching loss function OLP preferential with difficult sample knows another characteristic again to pedestrian and instructs
Practice, extracts the feature that can be used for pedestrian's identification mission again.
As shown in Fig. 2, the costing bio disturbance of OLP experienced three steps:
1), being used in newly added full articulamentum, pedestrian knows another characteristic again and its corresponding pedestrian's identity label ID is deposited
It stores up inside characteristic storage module;
2) reference sample and positive sample, are found in the pedestrian in full articulamentum again identification feature, from characteristic storage module
The middle label sample different from reference sample of finding is as negative sample;
3) positive sample and the corresponding feature of negative sample, searched out by two above step carries out loss function calculating, OLP
The calculating of loss function is as follows:
In formula,The feature of i-th of reference sample is represented,The feature of corresponding positive sample is represented,From network
It extracts,The feature for representing negative sample is extracted, n from characteristic storage modulejThe number of negative sample is represented, m is reference sample
Number, K be negative sample number, d () represent calculate two features between COS distance.
In this loss function, a sample is regarded as in each recommendation region generated by RPN network.
Gradient is calculated to loss function, available following formula:
In formula:
L=1,2 ..., K
The feature for representing negative sample is extracted, n from characteristic storage modulelRepresent the number of negative sample.
Using the back-propagation algorithm in neural network, present invention utilizes the modes of stochastic gradient descent SGD to update
Parameter in network.
The loss function of HEP is as follows:
HEP loss function is further learnt pedestrian in the way of classification and knows another characteristic again, and the present invention is according to data set
The identity ID of the pedestrian marked classifies come the area-of-interest generated to RPN, final to be divided into N+1 class altogether,
In, N represents the number of the identity for the pedestrian contained in data set, and the one kind added is then background classes.In each iteration, from
Middle selection C class (C≤N+1) carries out costing bio disturbance, it is assumed that the category set of C class composition is L, and the classification L being selected is by following three
A step is determined:
1), all ID existing in input picture are selected as classification to be selected, is put into L;
2) it, for each sample, choosesIn the sample nearest from positive sample, corresponded to
Classification be put into L;
3), if the class number in set L is randomly chosen other ID still less than C, and is stored in category set
In L;The then expression formula of HEP loss function are as follows:
In formula, m is the number of reference sample, and C is the class number chosen, and 1 () is indicated if the formula of bracket class is full
Foot, this result are 1, otherwise are 0;Label represents the label (its corresponding classification) of reference sample;Indicate what network was exported
The score for belonging to kth class of i-th of reference sample,Indicate point for belonging to jth class for i-th of reference sample that network is exported
Number.
Likewise, can use HEP damage using the back-propagation algorithm and stochastic gradient descent SGD algorithm of neural network
Function is lost to be updated the parameter of Faster-RCNN.
It is right respectively in loss functionWithDerivation has then for single sample:
BP is carried out to gradient reversely to return, and weight parameter is updated using stochastic gradient descent SGD, Ke Yigeng
The final argument of new network.
Under the collective effect of the relevant loss function of detection of OLP, HEP and Faster-RCNN itself, entirely
Faster-RCNN is able to training.
The test of Faster-RCNN after step 3, improvement
It inputs after a test picture, via trained parameter, available final detection block coordinate and its right
The feature of the detected pedestrian answered;Calculate pedestrian corresponding to detection block in different pictures again identification feature COS distance simultaneously
It is compared, the maximum two pedestrian detection frames of COS distance can determine whether from the same pedestrian.
Embodiment:
1, data set
Using CUHK-SYSU data set, a picture for sharing 18184 different scenes in data set.The street notebook data Ji You
The picture clapped on picture and film marks, and is relatively suitble to the training and survey of the detection of pedestrian and the identification mission again of pedestrian
Examination.
2, experimental setup
Training set has 11206 pictures, and the pedestrian of the different identity marked comprising 5532, test set has 6978 figures
Piece, the pedestrian of the different identity marked comprising 2900.
In the training process, the image that we input is to being to be matched based on 5532 pedestrian's identity being marked to picture
It is right, ultimately form 16000 images pair.
3, training test method
Training stage: will be trained in the pairs of input network of the image matched, and every two pairs of samples calculate primary ladder
The average value of degree carries out the parameter in a SGD update network.Terminate to obtain the final result of network after iteration 60000 times.
Test phase: we will test picture and input trained network model, detect the position of pedestrian and extraction pair
The feature answered, evaluation method are carried out according to the evaluation method of CUHK-SYSU, are calculated mAP (mean Average Precision)
With Top-1 index.
MAP index and AP hereafter, Recall index are recorded in The pascal visual object classes
(voc) challenge. (challenge of Pascal VOC object category) Everingham, M., Gool, L.V., Williams,
C.K.I.,Winn,J.,&Zisserman,A.(2010).International Journal of Computer Vision,
88(2),303-338.
Recognition accuracy compares
In order to verify effectiveness of the invention, the present embodiment combines different pedestrian detections and pedestrian's recognition methods conduct again
Comparison of the invention, there are four types of the pedestrian detections for comparing: CCF, ACF, Faster-RCNN (CNN), GT;
Existing pedestrian again recognition methods have it is following several:
1, three kinds of pedestrians identification feature extracting method DSIFT, BoW, LOMO and four kinds of characteristic measure methods again
Euclidean, KISSME, Cosine, XQDA combination;
2, two pedestrian detection and identifying system OIM and NPSM models again end to end.
Four kinds of pedestrian detection method foundations:
1, " Convolutional Channel features " (convolution channels feature) Yang, B., Yan, J., Lei,
Z. the CCF method recorded in, &Li, S.Z. (2015);
2, " Fast feature pyramids for object the detection " (swift nature for object detection
Pyramid) Dollar, P., Appel, R., Belongie, S., &Perona, P. (2014) .IEEE Transactions on
Pattern Analysis&Machine Intelligence, 36 (8), the ACF method recorded in 1532-45.;
3, Faster-RCNN (abbreviation CNN);
4, GT (the detection target manually extracted).
Three kinds of pedestrians identification feature extracting method foundation again:
1, " Unsupervised Salience Learning for Person Re-identification. " (is used for
The unsupervised significant inquiry learning that pedestrian identifies again) Zhao, Rui, W.Ouyang, and X.Wang.IEEE Conference on
Computer Vision and Pattern Recognition IEEE Computer Society,2013:3586-3593.
The DSIFT method of middle record;
2, " Scalable Person Re-identification:A Benchmark " (pedestrian of upgrading identify again appoint
Business: a new data set) Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., &Tian, Q. (2016)
.IEEE it is recorded in International Conference on Computer Vision (pp.1116-1124) .IEEE.
BoW method;
3、“Person re-identification by Local Maximal Occurrence
Representation and metric learning " (is indicated and the pedestrian side of identification again of metric learning based on local maxima
Method) Liao, S., Hu, Y., Zhu, X., &Li, S.Z. (2015) .IEEE Conference on Computer Vision and
The LOMO method recorded in Pattern Recognition (Vol.8, pp.2197-2206) .IEEE.
Four kinds of characteristic measure method foundations:
1, Euclidean (Euclidean distance);
2, " Large scale metric learning from equivalence constraints. " (based on etc.
The metric learning of valence constraint) Hirzer, M. (2012) .IEEE Conference on Computer Vision and
The KISSME method recorded in Pattern Recognition (pp.2288-2295) .IEEE Computer Society.;
3, Cosine (COS distance);
4、Liao,S.,Hu,Y.,Zhu,X.,&Li,S.Z.(2015).IEEE Conference on Computer
The XQDA method recorded in Vision and Pattern Recognition (Vol.8, pp.2197-2206) .IEEE.
Two pedestrian detection and identifying system foundations again end to end:
1、“Joint Detection and Identification Feature Learning for Person
Search " (while knowing another characteristic again with pedestrian for detecting to carry out pedestrian retrieval) .Xiao, T., Li, S., Wang, B.,
Lin,L.,&Wang,X.(2017).Computer Vision and Pattern Recognition(pp.3376-3385)
.IEEE. the OIM model recorded in;
2, " Neural person search machines " (pedestrian retrieval machine neural network based) .Liu, H.,
The NPSM recorded in Feng, J., Jie, Z., Jayashree, K., Zhao, B., &Qi, M., et al. (2017) .493-501.
Model.
Training test the results are shown in Table 1:
Table 1, the present invention are compared with other again recognition methods
Table 2, the present invention are compared with OIM detection effect
Method | AP (%) | Recall (%) |
OIM | 74.9 | 79.1 |
The present invention | 79.6 | 82.2 |
As can be seen from Table 1 and Table 2: of the invention (I-net) is expert at the effect obtained on personal data collection than existing pedestrian
It is good with the effect of recognition methods again to detect.
Table 3, several loss functions combined performance compare
Lose type | MAP (%) | Top-1 (%) |
On-line matching loss function | 73.6 | 76.2 |
On-line matching loss function+softmax | 79.0 | 81.2 |
On-line matching loss function+preferential the loss function of hardly possible sample | 79.5 | 81.5 |
The performance of the memory module storage number of features of table 4, OLP compares
As can be seen from tables 3 and 4 that the present invention can using On-line matching loss function+preferential loss function of hardly possible sample
To obtain better effect.
Claims (4)
1. a kind of processing method that the pedestrian detection based on deep neural network identifies again, in Faster-RCNN object detection net
On the basis of network, characterized in that further comprising the steps of:
Step 1 improves Faster-RCNN
1) Faster-RCNN, is added to an additional region recommendation network RPN, enables improved Faster-RCNN network
Enough while two pictures of input, and obtain the corresponding recommendation region of each picture;
2), in the full articulamentum for being finally added to 256 neurons of network for extracting the relevant feature of pedestrian's identity,
And add calculating of the characteristic storage module for loss function;
3) it, is added to On-line matching loss function OLP after the full articulamentum identified again for pedestrian and difficult sample preferentially damages
Function HEP is lost, is learnt for knowing another characteristic again to pedestrian;
The training of Faster-RCNN after step 2, improvement
Two picture inputs containing identical pedestrian are improved Faster-RCNN network, utilize the region recommendation network RPN of two-way
The recommendation frame proposals of two road networks is extracted respectively, is recycled and is recommended pool area layer ROIpooling that convolutional layer is above pushed away
The feature for recommending frame corresponding position is sent into full articulamentum, and full articulamentum extracts simultaneously to recommending frame to carry out further screening, amendment
Identification feature, loss function HEP addition On-line matching loss function OLP preferential with difficult sample are being used to extract pedestrian pedestrian again
Again behind the 256 full articulamentums of dimension of identification feature, the study of whole network is supervised;
The test of Faster-RCNN after step 3, improvement
A test picture is inputted into improved Faster-RCNN, is calculated, is obtained most using trained network parameter
Whole pedestrian detection result and pedestrian identifies required feature again.
2. the processing method that the pedestrian detection according to claim 1 based on deep neural network identifies again, characterized in that
The costing bio disturbance of On-line matching loss function OLP undergo the following three steps:
1), being used in newly added full articulamentum, pedestrian knows another characteristic again and its corresponding pedestrian's identity label ID storage is arrived
Inside characteristic storage module;
2) reference sample and positive sample, are found in the pedestrian in full articulamentum again identification feature, are sought from characteristic storage module
The sample for looking for label different from reference sample is as negative sample;
3) positive sample and the corresponding feature of negative sample, searched out by two above step carries out loss function calculating,
The calculating of OLP loss function is as follows:
In formula,The feature of i-th of reference sample is represented,The feature of corresponding positive sample is represented,It is extracted from network,The feature for representing negative sample is extracted, n from characteristic storage modulejThe number of negative sample is represented, m is of reference sample
Number, K are the number of negative sample, and d () represents the COS distance calculated between two features.
3. the processing method that the pedestrian detection according to claim 2 based on deep neural network identifies again, characterized in that
The parameter in network is updated using stochastic gradient descent SGD, it is as follows to calculate gradient formula to OLP loss function:
In formula:
The feature for representing negative sample is extracted, n from characteristic storage modulelRepresent the number of negative sample.
4. the processing method that the pedestrian detection according to claim 3 based on deep neural network identifies again, characterized in that
The identity ID mono- of pedestrian is divided into N+1 class, in each iteration, C class is selected to carry out costing bio disturbance, C class composition from N+1 class
Category set be L, the classification L being selected determined by following three steps:
1), all ID existing in input picture are selected as classification to be selected, is put into L;
2) it, for each sample, choosesIn the sample nearest from positive sample, by its corresponding class
L is not put into it;
3), if the class number in set L is randomly chosen other ID still less than C, and is stored in category set L;
The then expression formula of HEP loss function are as follows:
In formula, m is the number of reference sample, and C is the class number chosen, and 1 () is indicated if the formula of bracket class meets, this
As a result it is 1, otherwise is 0;Label represents the label of reference sample;Indicate belonging to for i-th of reference sample that network is exported
The score of kth class,Indicate the score for belonging to jth class for i-th of reference sample that network is exported.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810888879.6A CN109271852A (en) | 2018-08-07 | 2018-08-07 | A kind of processing method that the pedestrian detection based on deep neural network identifies again |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810888879.6A CN109271852A (en) | 2018-08-07 | 2018-08-07 | A kind of processing method that the pedestrian detection based on deep neural network identifies again |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109271852A true CN109271852A (en) | 2019-01-25 |
Family
ID=65153385
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810888879.6A Pending CN109271852A (en) | 2018-08-07 | 2018-08-07 | A kind of processing method that the pedestrian detection based on deep neural network identifies again |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109271852A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109886245A (en) * | 2019-03-02 | 2019-06-14 | 山东大学 | A kind of pedestrian detection recognition methods based on deep learning cascade neural network |
CN110097053A (en) * | 2019-04-24 | 2019-08-06 | 上海电力学院 | A kind of power equipment appearance defect inspection method based on improvement Faster-RCNN |
CN110503160A (en) * | 2019-08-28 | 2019-11-26 | 北京达佳互联信息技术有限公司 | Image-recognizing method, device, electronic equipment and storage medium |
CN111046724A (en) * | 2019-10-21 | 2020-04-21 | 武汉大学 | Pedestrian retrieval method based on area matching network |
CN111738081A (en) * | 2020-05-20 | 2020-10-02 | 杭州电子科技大学 | Deep neural network sonar target detection method difficult for sample retraining |
CN112907991A (en) * | 2021-02-03 | 2021-06-04 | 长安大学 | Traffic light signal time delay method, device, equipment and medium for courtesy pedestrians |
CN113505724A (en) * | 2021-07-23 | 2021-10-15 | 上海应用技术大学 | Traffic sign recognition model training method and system based on YOLOv4 |
US11321590B2 (en) | 2019-12-31 | 2022-05-03 | Industrial Technology Research Institute | Training method and system of objects detection model based on adaptive annotation design |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106845374A (en) * | 2017-01-06 | 2017-06-13 | 清华大学 | Pedestrian detection method and detection means based on deep learning |
CN107273872A (en) * | 2017-07-13 | 2017-10-20 | 北京大学深圳研究生院 | The depth discrimination net model methodology recognized again for pedestrian in image or video |
CN107330355A (en) * | 2017-05-11 | 2017-11-07 | 中山大学 | A kind of depth pedestrian based on positive sample Constraints of Equilibrium identification method again |
-
2018
- 2018-08-07 CN CN201810888879.6A patent/CN109271852A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106845374A (en) * | 2017-01-06 | 2017-06-13 | 清华大学 | Pedestrian detection method and detection means based on deep learning |
CN107330355A (en) * | 2017-05-11 | 2017-11-07 | 中山大学 | A kind of depth pedestrian based on positive sample Constraints of Equilibrium identification method again |
CN107273872A (en) * | 2017-07-13 | 2017-10-20 | 北京大学深圳研究生院 | The depth discrimination net model methodology recognized again for pedestrian in image or video |
Non-Patent Citations (2)
Title |
---|
SHAOQING REN ET AL: "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks", 《ARXIV:1506.01497V3》 * |
ZHENWEI HE ET AL: "End-to-End Detection and Re-identification Integrated Net for Person Search", 《ARXIV:1804.00376V1》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109886245A (en) * | 2019-03-02 | 2019-06-14 | 山东大学 | A kind of pedestrian detection recognition methods based on deep learning cascade neural network |
CN110097053A (en) * | 2019-04-24 | 2019-08-06 | 上海电力学院 | A kind of power equipment appearance defect inspection method based on improvement Faster-RCNN |
CN110503160A (en) * | 2019-08-28 | 2019-11-26 | 北京达佳互联信息技术有限公司 | Image-recognizing method, device, electronic equipment and storage medium |
CN111046724A (en) * | 2019-10-21 | 2020-04-21 | 武汉大学 | Pedestrian retrieval method based on area matching network |
CN111046724B (en) * | 2019-10-21 | 2021-09-14 | 武汉大学 | Pedestrian retrieval method based on area matching network |
US11321590B2 (en) | 2019-12-31 | 2022-05-03 | Industrial Technology Research Institute | Training method and system of objects detection model based on adaptive annotation design |
CN111738081A (en) * | 2020-05-20 | 2020-10-02 | 杭州电子科技大学 | Deep neural network sonar target detection method difficult for sample retraining |
CN112907991A (en) * | 2021-02-03 | 2021-06-04 | 长安大学 | Traffic light signal time delay method, device, equipment and medium for courtesy pedestrians |
CN113505724A (en) * | 2021-07-23 | 2021-10-15 | 上海应用技术大学 | Traffic sign recognition model training method and system based on YOLOv4 |
CN113505724B (en) * | 2021-07-23 | 2024-04-19 | 上海应用技术大学 | YOLOv 4-based traffic sign recognition model training method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109271852A (en) | A kind of processing method that the pedestrian detection based on deep neural network identifies again | |
CN109740413B (en) | Pedestrian re-identification method, device, computer equipment and computer storage medium | |
Ren et al. | A generic deep-learning-based approach for automated surface inspection | |
Pereira et al. | A deep learning-based approach for road pothole detection in timor leste | |
CN108520226B (en) | Pedestrian re-identification method based on body decomposition and significance detection | |
CN108197587A (en) | A kind of method that multi-modal recognition of face is carried out by face depth prediction | |
Aslam et al. | On the application of automated machine vision for leather defect inspection and grading: a survey | |
CN109816689A (en) | A kind of motion target tracking method that multilayer convolution feature adaptively merges | |
CN110991348B (en) | Face micro-expression detection method based on optical flow gradient amplitude characteristics | |
CN104504365A (en) | System and method for smiling face recognition in video sequence | |
CN109800624A (en) | A kind of multi-object tracking method identified again based on pedestrian | |
CN106846355A (en) | Method for tracking target and device based on lifting intuitionistic fuzzy tree | |
CN108090406A (en) | Face identification method and system | |
CN105389562A (en) | Secondary optimization method for monitoring video pedestrian re-identification result based on space-time constraint | |
CN109034035A (en) | Pedestrian's recognition methods again based on conspicuousness detection and Fusion Features | |
CN108764018A (en) | A kind of multitask vehicle based on convolutional neural networks recognition methods and device again | |
Pereira et al. | Classification of paved and unpaved road image using convolutional neural network for road condition inspection system | |
Zhang et al. | Identification of tomato leaf diseases based on multi-channel automatic orientation recurrent attention network | |
CN111539320B (en) | Multi-view gait recognition method and system based on mutual learning network strategy | |
CN116977937A (en) | Pedestrian re-identification method and system | |
CN109657586B (en) | Facial expression analysis method and system based on sequencing convolutional neural network | |
CN113723188A (en) | Dress uniform person identity verification method combining face and gait features | |
CN113674225A (en) | Power equipment fault detection method based on convolutional neural network | |
Aktas et al. | Multiple instance learning for CNN based fire detection and localization | |
CN109711232A (en) | Deep learning pedestrian recognition methods again based on multiple objective function |
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 |
Application publication date: 20190125 |
|
RJ01 | Rejection of invention patent application after publication |