CN110020688A - Pedestrian detection method is blocked based on deep learning - Google Patents

Pedestrian detection method is blocked based on deep learning Download PDF

Info

Publication number
CN110020688A
CN110020688A CN201910286482.4A CN201910286482A CN110020688A CN 110020688 A CN110020688 A CN 110020688A CN 201910286482 A CN201910286482 A CN 201910286482A CN 110020688 A CN110020688 A CN 110020688A
Authority
CN
China
Prior art keywords
feature
network
convolution
candidate region
pedestrian
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.)
Granted
Application number
CN201910286482.4A
Other languages
Chinese (zh)
Other versions
CN110020688B (en
Inventor
韩冰
王云浩
杨铮
朱考进
韩怡园
王平
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Original Assignee
Xidian University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xidian University filed Critical Xidian University
Priority to CN201910286482.4A priority Critical patent/CN110020688B/en
Publication of CN110020688A publication Critical patent/CN110020688A/en
Application granted granted Critical
Publication of CN110020688B publication Critical patent/CN110020688B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

Pedestrian detection method is blocked based on deep learning the invention discloses a kind of.Mainly solve the problems, such as that the prior art is poor to pedestrian detection effect is blocked.Its implementation is: reading pedestrian detection database data, extracts feature using VGG network；The feature that VGG network different layers extract is merged, obtains fusion feature, and using VGG network the last layer feature as non-fused feature；Mask network is constructed, fusion feature and non-fused feature are separately input in mask network, two kinds of convolution features for having missing are obtained；Building differentiates network, and two kinds of features that mask network is obtained are input to differentiation network, select more effective feature；Candidate region is obtained using the feature selected, and final testing result is obtained by candidate region.The present invention is improved to the detection effect for blocking pedestrian, be can be used for unmanned and auxiliary and is driven.

Description

Pedestrian detection method is blocked based on deep learning
Technical field
The invention belongs to technical field of image processing, in particular to a kind of for pedestrian detection method is blocked, and can be used for nothing People drives or auxiliary drives.
Background technique
Currently, mode identification technology and computer vision technique are very powerful and exceedingly arrogant, and sum of the pedestrian detection as image procossing One research field of machine vision, plays the important and pivotal role.The extensive use of machine vision has led new science and technology Trend.Therefore, more and more scholars have put among this strand of tide.This emerging technology revolution the whole world sweep across and Come, the eye of strategy is put into this block field by global scholar and businessman, strives that commanding elevation can be obtained.
In recent years, more and more universities, car manufacturer, internet giant, national scientific research institution are all set up specially Research center, strive mode identification technology and machine vision technique capable of being applied to industry, commercial field, create huge Value.And hot spot of the pedestrian detection as image procossing and machine vision the supreme arrogance of a person with great power, even more receive more extensive pass Note.Between in recent years, with artificial intelligence and internet explosive growth, European Union, which furnishes a huge amount of money for subsidize, has set up multiple pedestrian detection systems System；Pedestrian detection is applied to automobile assistant driving system by Honda Motors Co., to improve the safe maneuverability of automobile；Apple Pedestrian detection is applied to pilotless automobile already by the internets giant such as fruit, Google, FACEBOOK, by answering pedestrian detection Use intelligent auxiliary driving system for automobiles, Lai Yinling scientific and technological revolution next time.
Pedestrian detection becomes a very challenging research topic.Existing method includes special based on manual extraction The method of sign and method based on deep learning.
Method based on manual extraction feature has the fast development of more than ten years, and many scholars have done a large amount of work. 2005, Dalal et al. proposed the local variance of gradient orientation histogram characterization image, and has used linear supporting vector The classification of machine progress feature.2007, Dollar et al. was combined with local channel feature and the boosting algorithm of standard carries out Pedestrian detection.2013, R.Benenson et al. used local channel feature (ICF) to carry out pedestrian detection, and to influence ICF The many factors of effect have done detailed discussion.The same year, Prioletti et al. are produced using the cascade classifier based on Haar feature Raw pedestrian target region that may be present, and further confirm that in these regions whether deposit by the filter based on HOG feature In pedestrian target.2014, Nam et al. proposed to carry out the Gradient Features of pedestrian target and color characteristic at local decorrelation Reason enhances the classification capacity of boosted classifier.2017, J.Baekthe et al. proposed a kind of " additional kernel support vectors Machine (AKSVM) " is used as feature classifiers, and is optimized using genetic algorithm to AKSVM.However manual extraction characterization method Present people have much been not achieved for high-precision demand in accuracy rate.
With the extensive substantially enhancing increased with computing capability of training data, deep learning method is led in pedestrian detection Domain achieves success.2015, Yang et al. extracted the low-level feature in convolutional channel feature CCF, used enhancing forest model Pedestrian detection is realized as feature classifiers.In the same year, Cai et al. proposes that complicated perception cascades training CompACT, for integrating hand The feature that work extracts feature and CNN is extracted, has obtained a good balance between accuracy and speed.2016, Zhang etc. People obtains the convolution feature of pedestrian target using RPN, then realizes pedestrian detection using enhancing forest classified device.2018, You Et al. used the enhancing of simple three convolutional layers " aggregation channel characteristics ", and input figure is judged using adaboost classifier It seem no comprising pedestrian.However, above-mentioned deep learning method does not account for convolutional layer for different input picture expression effects Different problems do not account in actual conditions the problem of blocking pedestrian sample negligible amounts yet, cause the above method to screening The detection effect for keeping off pedestrian is poor.
Summary of the invention
It is an object of the invention to be directed to the deficiency of above-mentioned prior art, propose a kind of to block pedestrian based on deep learning Detection method, to improve to the detection effect for blocking pedestrian.
To achieve the above object, the invention proposes differentiation networks and mask network, implementation includes the following:
(1) pedestrian detection database data is read, the convolution feature of data is extracted using VGG convolutional neural networks, by VGG The feature that convolutional neural networks different layers extract is overlapped fusion, obtains fusion feature, and by VGG network the last layer feature As non-fused feature；
(2) mask network is constructed, fusion feature and non-fused feature are separately input in the mask network, obtain two kinds The convolution feature bad to pedestrian's expression effect；
(3) the differentiation network being made of mask network, RPN network and softmax classifier, two that (2) are obtained are constructed Kind of convolution feature is separately input to obtain two kinds in RPN network may be containing the candidate region of pedestrian target, by the candidate region It is input in softmax classifier, obtains two kinds of probability scores, the numerical value of both probability scores is between 0 to 1；According to Probability score selection output is for the more efficiently convolution feature of pedestrian target under blocking, when the probability that fusion feature obtains obtains Point be higher than non-fused feature obtain probability score when, export fusion feature, conversely, then exporting non-fused feature；
(4) it according to the feature exported in (3), obtains returning boundary and class probability:
4a) feature exported in (3) is input in RPN network, the candidate region of pedestrian target is obtained, candidate region It is mapped in the convolution characteristic layer of VGG convolutional neural networks, obtains each candidate region corresponding convolution in convolution characteristic layer Feature；
4b) the convolution feature that 4a) is obtained is input in two full articulamentums, thousands of class probabilities are obtained, wherein often One class probability has its corresponding recurrence boundary；
(5) according to 4b) obtained recurrence boundary and class probability, by loss function L to VGG convolutional neural networks, sentence Other network and RPN network are trained, and obtain final network model:
5a) subfunction L is lost using the class probability that loss function L includesclsWith recurrence marginal loss subfunction Lreg, meter It calculates class probability and loses subfunction LclsWith recurrence marginal loss subfunction Lreg；
Subfunction L 5b) is lost according to class probabilityclsWith recurrence marginal loss subfunction LregValue, loss is calculated Function L；
The value for 5c) reducing loss function L by progressive alternate is completed to VGG convolutional neural networks, differentiates network and RPN The training of network obtains final network model；
(6) image to be detected is input in final network model, the classification for obtaining thousands of image to be detected is general Rate, wherein each class probability has its corresponding recurrence boundary；Remain larger than the class probability of given threshold, and by these points The corresponding recurrence boundary of class probability is mapped in image to be detected, obtains one or more rectangle frames, as final detection As a result.
The present invention differentiates network and mask network due to constructing, and can choose out for the pedestrian target under blocking more Effective convolution feature, improves to the detection effect for blocking pedestrian.
Detailed description of the invention
Fig. 1 is implementation flow chart of the invention；
Fig. 2 is differentiation network and mask schematic network structure in the present invention；
Fig. 3 is with the present invention to the experimental result picture for blocking pedestrian in Caltech pedestrian detection database；
Fig. 4 is the experimental result picture with the present invention to all pedestrians of Caltech pedestrian detection database.
Specific embodiment
The embodiment of the present invention and effect are further described below in conjunction with attached drawing.
Referring to Fig.1, specific implementation step of the invention is as follows:
Step 1, fusion feature and non-fused feature are obtained.
1.1) pedestrian detection database data is read, the convolution feature of data is extracted using VGG convolutional neural networks:
Convolutional calculation 1.1a) is carried out to input data using convolution kernel, obtains first convolution feature in neural network Figure；
Convolutional calculation 1.1b) is carried out to current convolution characteristic pattern using convolution kernel, it is special to obtain the next convolution of neural network Sign figure；
1.1c) repeat step 1.1b), 16 convolution characteristic patterns are calculated, in the 16th convolution characteristic pattern Data are the convolution feature finally extracted；
1.2) the convolution feature that VGG convolutional neural networks different layers extract is overlapped fusion, obtains fusion feature, and Using VGG network the last layer feature as non-fused feature.
Step 2, mask network is constructed, the convolution feature bad to pedestrian's expression effect is obtained.
2.1) mask network is set to be of five storeys altogether, in which:
First layer is convolutional layer, for carrying out convolution operation to having a size of w × h × c input feature vector, is obtained having a size of w The characteristic pattern of × h × c/16；
The second layer is pond layer, and the characteristic pattern for obtaining to first layer convolution carries out the pondization that step-length is 4 and operates, and is obtained Width is the characteristic pattern of w/4, a height of h/4；
Third layer is up-sampling, and the characteristic pattern for obtaining to second of pond up-samples, and obtaining width is w, a height of h Characteristic pattern；
4th layer is fused layer, and characteristic pattern and input feature vector figure for third layer to be up-sampled are merged, obtained To the characteristic pattern for being 17/16 × c having a size of depth；
Layer 5 is convolutional layer, carries out convolution operation for merging obtained characteristic pattern to the 4th layer, obtain having a size of w × The characteristic pattern of h × c；
2.2) fusion feature that step 1 obtains and non-fused feature are separately input in mask network, obtain two kinds it is right The bad convolution feature of pedestrian's expression effect.
Step 3, building differentiates network, selection output feature.
3.1) the mask network for constructing step 2 is in parallel with existing RPN network and softmax classifier, constitutes and differentiates Network, as shown in Figure 2, in which:
RPN network is used then using a picture of any scale as input, exports a series of rectangle candidate region, Each candidate region has the network of a classification score；
Softmax classifier, input is a vector, and output is normalized class probability；,
3.2) two kinds of convolution features for obtaining step 2.2) are separately input in RPN network, and obtaining two kinds may contain The candidate region is input in softmax classifier by the candidate region of pedestrian target, obtains two kinds of probability scores, both The numerical value of probability score is between 0 to 1；
3.3) according to probability score selection output for the more efficiently convolution feature of pedestrian target under blocking: working as fusion When the probability score that feature obtains is higher than the probability score that non-fused feature obtains, fusion feature is exported, conversely, then exporting non-melt Close feature.
Step 4, class probability is obtained according to the feature of differentiation network output and returns boundary.
4.1) feature for differentiating network output is input in RPN network, obtains the candidate region of pedestrian target, candidate It is corresponding in convolution characteristic layer to obtain each candidate region into the convolution characteristic layer of VGG convolutional neural networks for area maps Convolution feature；
4.2) full articulamentum is connected behind RPN network, 4.1) the convolution feature that middle mapping obtains is input to full connection In layer, each node of full articulamentum is connected with all nodes of convolution feature of input, then the spy that will be extracted before Sign integrates, and obtains the two-dimensional array that length is 3000-4000, and each of two-dimensional array value represents each candidate regions Domain belongs to the probability of pedestrian and background, obtains thousands of class probabilities by the thousands of values of array, wherein each class probability There is its corresponding recurrence boundary.
Step 5, using class probability and recurrence feature modeling loss function, by loss function to VGG convolutional Neural net Network differentiates that network and RPN network are trained, and obtains final network model.
5.1) the loss subfunction L that loss function L includes presentation class probability is setclsWith loss for indicating recurrence boundary Function Lreg；
5.2) it is calculate by the following formula the loss subfunction L of class probabilitycls:
Wherein, i is the index of candidate region, piThe class probability of a pedestrian whether is represented for each candidate region, For the true tag of candidate region, if in candidate region being pedestrian,It is 1, otherwise,It is 0；
5.3) it is calculate by the following formula the loss subfunction L for returning boundaryreg:
Wherein, i is the index of candidate region, tiFor the recurrence boundary of candidate region,For the true seat of pedestrian region Mark；
5.4) according to above-mentioned two sub- LclsAnd LregValue, calculate loss function L:
Wherein, i is the index of candidate region, piThe probability of a pedestrian whether is represented for each candidate region,To wait The true tag along sort of favored area, if in candidate region being pedestrian,It is 1, otherwiseIt is 0；tiFor the seat of candidate region Mark,For the true coordinate of pedestrian region, NclsAnd NregFor the different normalization coefficient of two values, NclsValue is 256, NregValue is that 2400, λ is coefficient of balance；
5.5) value for reducing loss function L by progressive alternate is completed to VGG convolutional neural networks, differentiates network and RPN The training of network obtains final network model.
Step 6, testing result is obtained.
6.1) image to be detected is input in final network model, the classification for obtaining thousands of image to be detected is general Rate, wherein each class probability has its corresponding recurrence boundary；
6.2) threshold value is set as 0.5, remains larger than the class probability of the threshold value, and by pair of the class probability of reservation Boundary should be returned to be mapped in image to be detected, obtain one or more rectangle frames, as final testing result.
Effect of the invention is further described below with reference to emulation experiment.
1. simulated conditions:
On hardware facility, the high-performance calculation equipped with the smooth X video card of an I7-5930K processor and four pieces of Thailands is used Machine.
Experiment is assessed using Caltech pedestrian detection database, which is to advise at present The biggish pedestrian's database of mould, is shot using vehicle-mounted camera, is labelled with about 250000 frame images, shares 350000 rectangles Frame, including 2300 pedestrians.
Emulation experiment is enterprising in Caltech pedestrian detection database using the present invention and existing three kinds of pedestrian detection methods Capable comparative experiments, wherein the first existing method method is the convolution channel characteristics method CCF for being published in ICCV2015, second Kind existing method is the Area generation cascade enhancing forest method RPN+BF for being published in ECCV2016, the third existing method It is the characteristic binding learning method UDN+ for being published in TPAMI2017.
2. emulation content:
Emulation experiment 1: the row for being 40% to 80% to coverage extent in database with the present invention and existing three kinds of methods People detects, and obtains MR-FPPI curve, as shown in figure 3, wherein ordinate is Loss Rate MR, Loss Rate is that positive sample is wrong Erroneous judgement Wei not the number of negative sample and the ratio of whole positive sample numbers；Abscissa is wrong positive sample number in every image FPPI, wherein wrong positive sample refers to that testing result is pedestrian, actually not some samples of pedestrian.
As seen from Figure 3, the testing result for the pedestrian target that the present invention is 40% to 80% to coverage extent is better than other Three kinds of methods, demonstrating the present invention has good detection effect to blocking pedestrian.
Emulation experiment 2: with the present invention and existing three kinds of methods to pedestrian all in Caltech pedestrian detection database Target is detected, and MR-FPPI curve, such as Fig. 4 are obtained.
From fig. 4, it can be seen that the present invention is better than other three kinds of methods to the testing result of all pedestrians.Emulation experiment 2 Demonstrating the present invention equally has good detection effect to all pedestrians.
It is above-mentioned simulation results show correctness of the invention, validity and reliability.

Claims (7)

1. a kind of block pedestrian detection method based on deep learning characterized by comprising
(1) pedestrian detection database data is read, the convolution feature of data is extracted using VGG convolutional neural networks, by VGG convolution The feature that neural network different layers extract is overlapped fusion, obtains fusion feature, and using VGG network the last layer feature as Non-fused feature；
(2) mask network is constructed, fusion feature and non-fused feature are separately input in the mask network, obtain two kinds to row The bad convolution feature of people's expression effect；
(3) the differentiation network being made of mask network, RPN network and softmax classifier is constructed, two kinds of volumes that (2) are obtained Product feature is separately input to obtain two kinds in RPN network may to input the candidate region containing the candidate region of pedestrian target Into softmax classifier, two kinds of probability scores are obtained, the numerical value of both probability scores is between 0 to 1；According to probability Component selections output is obtained for the more efficiently convolution feature of pedestrian target under blocking, when the probability score that fusion feature obtains is high When the probability score that non-fused feature obtains, fusion feature is exported, conversely, then exporting non-fused feature；
(4) it according to the feature exported in (3), obtains returning boundary and class probability:
4a) feature exported in (3) is input in RPN network, the candidate region of pedestrian target is obtained, candidate region is mapped Into the convolution characteristic layer of VGG convolutional neural networks, each candidate region corresponding convolution feature in convolution characteristic layer is obtained；
4b) the convolution feature that 4a) is obtained is input in two full articulamentums, obtains thousands of class probabilities, wherein each Class probability has its corresponding recurrence boundary；
(5) according to 4b) obtained recurrence boundary and class probability, by loss function L to VGG convolutional neural networks, differentiate net Network and RPN network are trained, and obtain final network model:
5a) subfunction L is lost using the class probability that loss function L includesclsWith recurrence marginal loss subfunction Lreg, calculate and divide Class probability loses subfunction LclsWith recurrence marginal loss subfunction Lreg；
Subfunction L 5b) is lost according to class probabilityclsWith recurrence marginal loss subfunction LregValue, loss function is calculated L；
The value for 5c) reducing loss function L by progressive alternate is completed to VGG convolutional neural networks, differentiates network and RPN network Training, obtain final network model；
(6) image to be detected is input in final network model, obtains the class probability of thousands of image to be detected, In each class probability have its corresponding recurrence boundary；The class probability of given threshold is remained larger than, and these classification are general The corresponding recurrence boundary of rate is mapped in image to be detected, obtains one or more rectangle frames, as final testing result.
2. according to the method described in claim 1, the convolution feature of data is wherein extracted in (1) using VGG convolutional neural networks, It is to be extracted as follows by VGG convolutional neural networks by the convolution kernel of 3 × 3 matrixes:
Convolutional calculation 1a) is carried out to input data using convolution kernel, obtains first convolution characteristic pattern in neural network；
Convolutional calculation 1b) is carried out to current convolution characteristic pattern using convolution kernel, obtains the next convolution characteristic pattern of neural network；
1c) repeat step 1b), 16 convolution characteristic patterns are obtained, the data in the 16th convolution characteristic pattern are finally to mention The convolution feature taken out.
3. according to the method described in claim 1, the mask network wherein constructed in (2), structure are as follows:
First layer is convolutional layer, for carrying out convolution operation to having a size of w × h × c input feature vector figure, is obtained having a size of w × h The characteristic pattern of × c/16；
The second layer is pond layer, and the characteristic pattern for obtaining to first layer convolution carries out the pondization that step-length is 4 and operates, and obtaining width is The characteristic pattern of w/4, a height of h/4；
Third layer is up-sampling, and the characteristic pattern for obtaining to second of pond up-samples, and obtains the spy that width is w, a height of h Sign figure；
4th layer is fused layer, and characteristic pattern and input feature vector figure for third layer to be up-sampled are merged, and obtains ruler Very little is the characteristic pattern that depth is 17/16 × c；
Layer 5 is convolutional layer, carries out convolution operation for merging obtained characteristic pattern to the 4th layer, obtains having a size of w × h × c Characteristic pattern.
4. according to the method described in claim 1, wherein 4b) in the convolution feature for obtaining 4a) be input to two full connections In layer, thousands of class probabilities are obtained, are accomplished by
Classified using full articulamentum, i.e., each node of full articulamentum be connected with all nodes of input feature vector, The characteristic synthetic that will be extracted before again obtains the two-dimensional array that length is 3000-4000, each in the two-dimensional array A value represents the probability that each candidate region belongs to pedestrian and background, obtains thousands of class probabilities by the thousands of values of array.
5. according to the method described in claim 1, wherein step 5a) in calculate class probability and lose subfunction Lcls, pass through following formula It calculates:
Wherein, i is the index of candidate region, piThe detection probability of a pedestrian whether is represented for each candidate region,To wait The true tag of favored area, if in candidate region being pedestrian,It is 1, otherwise,It is 0.
6. according to the method described in claim 1, wherein step 5a) in calculate and return marginal loss subfunction Lreg, pass through following formula It calculates:
Wherein, i is the index of candidate region, tiFor the coordinate of candidate region,For the true coordinate of pedestrian region.
7. according to the method described in claim 1, wherein step 5b) in subfunction L lost according to class probabilityclsWith recurrence side Lose subfunction L in boundaryregValue, calculate loss function L, be calculate by the following formula:
Wherein, i is the index of candidate region, piThe probability of a pedestrian whether is represented for each candidate region,For candidate region True tag along sort, if in candidate region being pedestrian,It is 1, otherwiseIt is 0；tiFor the coordinate of candidate region,For The true coordinate of pedestrian region, NclsAnd NregFor the different normalization coefficient of two values, NclsValue is 256, NregIt takes Value is that 2400, λ is coefficient of balance.
CN201910286482.4A 2019-04-10 2019-04-10 Shielded pedestrian detection method based on deep learning Active CN110020688B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910286482.4A CN110020688B (en) 2019-04-10 2019-04-10 Shielded pedestrian detection method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910286482.4A CN110020688B (en) 2019-04-10 2019-04-10 Shielded pedestrian detection method based on deep learning

Publications (2)

Publication Number Publication Date
CN110020688A true CN110020688A (en) 2019-07-16
CN110020688B CN110020688B (en) 2022-12-06

Family

ID=67191051

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910286482.4A Active CN110020688B (en) 2019-04-10 2019-04-10 Shielded pedestrian detection method based on deep learning

Country Status (1)

Country Link
CN (1) CN110020688B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111046956A (en) * 2019-12-13 2020-04-21 苏州科达科技股份有限公司 Occlusion image detection method and device, electronic equipment and storage medium
CN111144203A (en) * 2019-11-19 2020-05-12 浙江工商大学 Pedestrian shielding detection method based on deep learning

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018214195A1 (en) * 2017-05-25 2018-11-29 中国矿业大学 Remote sensing imaging bridge detection method based on convolutional neural network
CN108960074A (en) * 2018-06-07 2018-12-07 西安电子科技大学 Small size pedestrian target detection method based on deep learning
CN109063559A (en) * 2018-06-28 2018-12-21 东南大学 A kind of pedestrian detection method returned based on improvement region

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018214195A1 (en) * 2017-05-25 2018-11-29 中国矿业大学 Remote sensing imaging bridge detection method based on convolutional neural network
CN108960074A (en) * 2018-06-07 2018-12-07 西安电子科技大学 Small size pedestrian target detection method based on deep learning
CN109063559A (en) * 2018-06-28 2018-12-21 东南大学 A kind of pedestrian detection method returned based on improvement region

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111144203A (en) * 2019-11-19 2020-05-12 浙江工商大学 Pedestrian shielding detection method based on deep learning
CN111144203B (en) * 2019-11-19 2023-06-16 浙江工商大学 Pedestrian shielding detection method based on deep learning
CN111046956A (en) * 2019-12-13 2020-04-21 苏州科达科技股份有限公司 Occlusion image detection method and device, electronic equipment and storage medium

Also Published As

Publication number Publication date
CN110020688B (en) 2022-12-06

Similar Documents

Publication Publication Date Title
CN110348384B (en) Small target vehicle attribute identification method based on feature fusion
CN109859190B (en) Target area detection method based on deep learning
CN110163187B (en) F-RCNN-based remote traffic sign detection and identification method
CN111898406B (en) Face detection method based on focus loss and multitask cascade
CN112801169B (en) Camouflage target detection method, system, device and storage medium based on improved YOLO algorithm
CN108509978A (en) The multi-class targets detection method and model of multi-stage characteristics fusion based on CNN
CN109543606A (en) A kind of face identification method that attention mechanism is added
CN110175613A (en) Street view image semantic segmentation method based on Analysis On Multi-scale Features and codec models
CN109934293A (en) Image-recognizing method, device, medium and obscure perception convolutional neural networks
CN108764063A (en) A kind of pyramidal remote sensing image time critical target identifying system of feature based and method
CN114758288B (en) Power distribution network engineering safety control detection method and device
CN112487862B (en) Garage pedestrian detection method based on improved EfficientDet model
CN112418236A (en) Automobile drivable area planning method based on multitask neural network
CN109492596B (en) Pedestrian detection method and system based on K-means clustering and regional recommendation network
CN113034506B (en) Remote sensing image semantic segmentation method and device, computer equipment and storage medium
CN109784197A (en) Pedestrian's recognition methods again based on hole convolution Yu attention study mechanism
CN114926747A (en) Remote sensing image directional target detection method based on multi-feature aggregation and interaction
CN109670555A (en) Instance-level pedestrian detection and pedestrian's weight identifying system based on deep learning
CN111652273A (en) Deep learning-based RGB-D image classification method
CN111739037B (en) Semantic segmentation method for indoor scene RGB-D image
CN117079098A (en) Space small target detection method based on position coding
CN112861970A (en) Fine-grained image classification method based on feature fusion
CN115909078A (en) Ship classification method based on HRRP and SAR data feature level fusion
CN110020688A (en) Pedestrian detection method is blocked based on deep learning
CN115346071A (en) Image classification method and system for high-confidence local feature and global feature 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
GR01 Patent grant
GR01 Patent grant