CN106845387B - Pedestrian detection method based on self-learning - Google Patents

Pedestrian detection method based on self-learning Download PDF

Info

Publication number
CN106845387B
CN106845387B CN201710033677.9A CN201710033677A CN106845387B CN 106845387 B CN106845387 B CN 106845387B CN 201710033677 A CN201710033677 A CN 201710033677A CN 106845387 B CN106845387 B CN 106845387B
Authority
CN
China
Prior art keywords
classifier
pedestrian
training
feature
mixture model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
CN201710033677.9A
Other languages
Chinese (zh)
Other versions
CN106845387A (en
Inventor
施培蓓
曹风云
胡玉娟
杨雪洁
王璐
钱言玉
王筱薇倩
张娜
谢超
吴友情
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.)
Hefei Normal University
Original Assignee
Hefei Normal 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 Hefei Normal University filed Critical Hefei Normal University
Priority to CN201710033677.9A priority Critical patent/CN106845387B/en
Publication of CN106845387A publication Critical patent/CN106845387A/en
Application granted granted Critical
Publication of CN106845387B publication Critical patent/CN106845387B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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
    • G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
    • 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
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a pedestrian detection method based on self-learning, which comprises the following specific steps of: firstly, training an AdaBoost-based cascade classifier as an offline classifier, simultaneously training a Gaussian mixture model by using a group of public pedestrian photos, adopting HOG (histogram of oriented gradient) features and position information for feature coding, then adopting a low-threshold offline classifier to detect pedestrians in a specific scene, outputting confidence scores of candidate objects, then selecting high confidence scores as positive samples, taking low confidence scores as negative samples, re-representing the candidate detection objects by using the Gaussian mixture model, finally training a discriminative pedestrian classifier on line by using an SVM (support vector machine) classifier, re-predicting the candidate objects and outputting probability estimation. The pedestrian detection method solves the problem that the traditional pedestrian detection method cannot be self-adaptive to a specific scene, has a certain promotion effect on the pedestrian detection technology under the specific scene, and is remarkably improved in the aspect of identification rate compared with the traditional pedestrian detection method.

Description

Pedestrian detection method based on self-learning
Technical Field
The invention relates to a pedestrian detection method in the field of intelligent transportation, in particular to a self-learning pedestrian detection method based on a specific scene.
Background
The problem of road traffic safety has seriously influenced economic development and social construction, and the reduction of the occurrence of road traffic accidents and casualties is an important matter of the problem of the relation of civilians. The problem of road traffic safety is influenced by a plurality of factors such as pedestrians, vehicles, roads and the like, and as the pedestrians are main participants and weak people of road traffic, ensuring the safety of the pedestrians becomes a key of the problem of road traffic safety, and is also an important task in the field of intelligent traffic systems.
The pedestrian detection method is a core support technology of an intelligent traffic system, and has profound influence on guaranteeing pedestrian safety and reducing life and property loss of people. In practical application, the vehicle-mounted pedestrian detection system is required to be capable of adapting to pedestrian detection in different scenes. At present, researchers and automobile manufacturers at home and abroad have recognized the important economic value and research significance of the application of the pedestrian detection system, and have proposed to definitely start to develop the research and application of automobile automatic driving, but have technical defects and potential safety hazards.
The pedestrian detection method mainly comprises two main categories of image processing and machine learning. The classification method in machine learning is a pedestrian detection technology which is adopted at present. The core technology of the pedestrian detection system based on classification is feature extraction and classifier design, and the basic idea is to train a complete classifier by using a large amount of training data and then detect pedestrians for test data. The pedestrian detection method based on classification is successfully applied in a fixed scene, and belongs to an off-line training mode. If the training set and the test set are from different sources, the classification performance of the offline-trained classifier is greatly reduced in a new scene due to the fact that the pedestrian features and the postures are greatly changed and the mismatching factors of the vision, the illumination, the background, the resolution ratio and the like of the scene are added. Experimental results of the existing dozens of pedestrian classifiers show that the classifier trained on the INRIA data set is directly used for pedestrian detection under other different scenes, and the omission ratio is improved by 20% to 50%. Therefore, it is difficult to train a general detector to be suitable for all pedestrian scenarios.
The classifier can be retrained for pedestrian detection in a specific scene, but the complex cost brought by re-standard samples is difficult to bear. The classifier requires a large number of training samples and training time to ensure good detection performance in the detection phase. In addition, the trained classifier parameters are already determined, and the environment of a new scene is difficult to adapt. Therefore, the pedestrian detection classification method for designing the self-adaptive scene has certain research and application values.
Disclosure of Invention
The purpose of the invention is as follows: in order to overcome the defect that self-adaption to a specific task cannot be achieved in the prior art, a self-learning-based pedestrian detection method is provided, any offline-trained pedestrian classifier can adapt to a specific scene, and a good recognition rate is obtained.
The technical scheme is as follows: in order to achieve the purpose, the invention provides a pedestrian detection method based on self-learning, which comprises the following specific steps: firstly, training an AdaBoost-based cascade classifier as an offline classifier, simultaneously training a Gaussian mixture model by using a group of public pedestrian photos, adopting HOG (histogram of oriented gradient) features and position information for feature coding, then adopting a low-threshold offline classifier to detect pedestrians in a specific scene, outputting confidence scores of candidate objects, then selecting high confidence scores as positive samples, taking low confidence scores as negative samples, re-representing the candidate detection objects by using the Gaussian mixture model, finally training a discriminative pedestrian classifier on line by using an SVM (support vector machine) classifier, re-predicting the candidate objects and outputting probability estimation.
The offline classifier in the above step: training data come from any pedestrian data set, and feature extraction adopts LBP features;
the Gaussian mixture model is as follows: training a Gaussian mixture model by adopting an INRIA data set in an off-line manner, learning parameters of the Gaussian mixture model by adopting an EM (effective electromagnetic modeling) algorithm, and for each pedestrian sample, extracting HOG (histogram of oriented gradient) features and position information of each image block on a multi-scale image to form a group of feature vectors of the pedestrian sample;
screening the candidate objects: setting a low threshold value of an off-line classifier, performing pedestrian detection on a specific scene by using the off-line classifier to obtain candidate detection objects and confidence scores thereof, setting two threshold values to screen all candidate objects, and obtaining a positive and negative sample set;
the online classifier: and for each pedestrian sample, regenerating a new feature representation by using a Gaussian mixture model, then training an SVM classifier on line, finally detecting the candidate object of the specific scene again by using the SVM classifier, and outputting probability estimation.
Further, the offline classifier is trained by adopting a cascade classifier in OpenCV, and comprises two parts, namely a data set preparation part and a training program operation part, wherein the data set preparation part is training data: creating a positive sample set by using opencv _ createsamples, and manually preparing a large number of negative sample pictures; the training program is run to train the cascade classifier: setting the feature type as LBP, and training a cascade classifier by adopting opencv _ traincacade.
Further, the gaussian mixture model includes two parts, namely feature coding and GMM training:
feature coding: for each pedestrian picture in the INRIA data set, firstly, three layers of Gaussian pyramids are constructed, and then overlapped image blocks are extracted from each layer of image pyramids. Assume that each pedestrian picture comprises N image blocks
Figure BDA0001212440140000021
Extracting HOG feature HOG of each image blockpiAnd its location information lpi＝[xy]TFinally, the feature of each image block is coded as fpi＝[hogpi T,lpi T]TAll image blocks constitute pedestrian sample features
Figure BDA0001212440140000022
And GMM training: an off-line trained Gaussian mixture model can be expressed as
Figure BDA0001212440140000023
Wherein K is the number of Gaussian mixture components,
Figure BDA0001212440140000024
i is a unit matrix of the image data,
Figure BDA0001212440140000025
is the mixing weight of the kth gaussian component,
Figure BDA0001212440140000026
is a Gaussian distribution with a mean value of μkAnd the variance is
Figure BDA0001212440140000027
f is the pedestrian sample feature in the feature encoding portion;
a set of pedestrian sample characteristics χ ═ f given in the INRIA dataset1,f2,...,fMAnd learning the parameter theta of the GMM by using the likelihood function of the characteristics of the maximum training set learned by the EM algorithm, wherein the parameter theta is expressed as
Figure BDA0001212440140000031
The EM algorithm is to calculate the solution by using two steps alternately, wherein the first step expectation (E) is used for calculating the expectation value of the log likelihood, and the second step maximization (M) is used for updating the parameters to find the parameter value of the maximization likelihood expectation, which is as follows:
e, step E:
Figure BDA0001212440140000032
Figure BDA0001212440140000033
Figure BDA0001212440140000034
wherein the content of the first and second substances,
Figure BDA0001212440140000035
is the kth Gaussian component generation feature fiA posterior probability of (d);
and M: updating the parameter Θ
Figure BDA0001212440140000036
Figure BDA0001212440140000037
Figure BDA0001212440140000038
Further, the specific method for screening the candidate objects comprises the following steps: the method comprises the steps of utilizing an offline-trained cascade classifier to detect pedestrians in a specific scene, setting a low detection threshold in order to obtain more detection objects, and ensuring that all pedestrian images are obtained, wherein a large number of false alarms exist, and all candidate objects in the specific scene are assumed to be T ═ T { (T) } TiI 1.. M }, each candidate outputting a corresponding confidence score S ═ S ·iI 1., M }, sorting the confidence scores in descending order, setting two thresholds λhAnd λlRepresenting positive and negative examples with high and low confidence scores, respectively, H and N are defined as follows:
H＝{ti:si＞λh,i＝1,...,M}
N＝{ti:si＜λl,i＝1,...,M}
we expect a balanced set of data H and N, so let C ═ min (| H |, | N |) then
H＝{t1,...,tC}
N＝{tM-C+1,...,tM}
C＜M/2
Further, the specific method of the online classifier is as follows: aiming at positive and negative sample sets H and N, firstly, expressing the positive and negative samples again by using an off-line trained Gaussian mixture model, then, training a classifier on line by using a LibSVM, and finally, using the on-line classifier to perform on-line classification on a candidate object tiIs a positive probability estimate. The method comprises the following specific steps:
step 1) characterization: for a single pedestrian sample feature is expressed as
Figure BDA0001212440140000041
Performing feature representation by using a Gaussian mixture model, wherein each Gaussian component is derived from a feature vector
Figure BDA0001212440140000042
Select a corresponding feature
Figure BDA0001212440140000046
Figure BDA0001212440140000043
For a single pedestrian sample picture, K features are selected by K Gaussian components, and the single pedestrian sample feature is re-expressed as [ f [ ]g1,fg2,...,fgK]And only the HOG characteristic is reserved for the final pedestrian sample characteristic, the position information of the image block is removed, and the final pedestrian sample characteristic is expressed as [ HOGg1,hogg2,...,hoggK]；
Step 2) training an SVM classifier: the SVM classifier is represented as
Figure BDA0001212440140000044
Wherein the kernel function adopts a Gaussian RBF kernel with the formula of
Figure BDA0001212440140000045
Step 3) prediction of an SVM classifier: using the classifier trained in the step 2) to perform all the candidate objects T ═ TiAnd predicting the I1, the other words, M, and outputting a probability estimation result.
Has the advantages that: compared with the prior art, the invention provides the pedestrian detection method based on self-learning, solves the problem that the traditional pedestrian detection method cannot be self-adapted to the specific scene, can change the common pedestrian classifier to be adapted to the specific scene, re-represents the candidate object of the specific scene through the off-line trained Gaussian mixture model, and has certain promotion effect on the pedestrian detection technology under the specific scene.
Drawings
FIG. 1 is a schematic structural view of the present invention;
FIG. 2 is a flow diagram of offline classifier training;
FIG. 3 is a flow chart of Gaussian mixture model training;
FIG. 4 is a flow chart of candidate screening;
FIG. 5 is a flow chart of classifier training.
Detailed Description
The present invention is further illustrated by the following figures and specific examples, which are to be understood as illustrative only and not as limiting the scope of the invention, which is to be given the full breadth of the appended claims and any and all equivalent modifications thereof which may occur to those skilled in the art upon reading the present specification.
Example 1:
as shown in fig. 1, the invention provides a pedestrian detection method based on self-learning, which comprises the following specific steps: firstly, training an AdaBoost-based cascade classifier as an offline classifier, simultaneously training a Gaussian mixture model by using a group of public pedestrian photos, adopting HOG (histogram of oriented gradient) features and position information for feature coding, then adopting a low-threshold offline classifier to detect pedestrians in a specific scene, outputting confidence scores of candidate objects, then selecting high confidence scores as positive samples, taking low confidence scores as negative samples, re-representing the candidate detection objects by using the Gaussian mixture model, finally training a discriminative pedestrian classifier on line by using an SVM (support vector machine) classifier, re-predicting the candidate objects and outputting probability estimation.
The offline classifier in the above step: training data come from any pedestrian data set, and feature extraction adopts LBP features;
the Gaussian mixture model is as follows: training a Gaussian mixture model by adopting an INRIA data set in an off-line manner, learning parameters of the Gaussian mixture model by adopting an EM (effective electromagnetic modeling) algorithm, and for each pedestrian sample, extracting HOG (histogram of oriented gradient) features and position information of each image block on a multi-scale image to form a group of feature vectors of the pedestrian sample;
screening the candidate objects: setting a low threshold value of an off-line classifier, performing pedestrian detection on a specific scene by using the off-line classifier to obtain candidate detection objects and confidence scores thereof, setting two threshold values to screen all candidate objects, and obtaining a positive and negative sample set;
the online classifier: and for each pedestrian sample, regenerating a new feature representation by using a Gaussian mixture model, then training an SVM classifier on line, finally detecting the candidate object of the specific scene again by using the SVM classifier, and outputting probability estimation.
Example 2:
as shown in fig. 2, the offline classifier is trained by using a cascade classifier in OpenCV, and includes two parts, namely, a data set preparation part and a training program running part, where the data set preparation part is training data: creating a positive sample set by using opencv _ createsamples, and manually preparing a large number of negative sample pictures; the training program is run to train the cascade classifier: setting the feature type as LBP, and considering that the training and detection speed of the LBP feature is several times faster than that of the Haar feature, adopting the LBP feature and adopting opencv _ traincascade to train the cascade classifier.
Example 3:
as shown in fig. 3, the gaussian mixture model includes two parts, namely feature coding and GMM training:
feature coding: for each pedestrian picture in the INRIA data set, firstly, three layers of Gaussian pyramids are constructed, and then overlapped image blocks are extracted from each layer of image pyramids. Assume that each pedestrian picture comprises N image blocks
Figure BDA0001212440140000061
Extracting HOG feature HOG of each image blockpiAnd its location information lpi＝[xy]TFinally, the feature of each image block is coded as fpi＝[hogpi T,lpi T]TAll image blocks constitute pedestrian sample features
Figure BDA0001212440140000062
And GMM training: an off-line trained Gaussian mixture model can be expressed as
Figure BDA0001212440140000063
Wherein K is the number of Gaussian mixture components,
Figure BDA0001212440140000064
i is a unit matrix of the image data,
Figure BDA0001212440140000065
is the mixing weight of the kth gaussian component,
Figure BDA0001212440140000066
is a Gaussian distribution with a mean value of μkAnd the variance is
Figure BDA0001212440140000067
f is the pedestrian sample feature in the feature encoding portion;
a set of pedestrian sample characteristics χ ═ f given in the INRIA dataset1,f2,...,fMAnd learning the parameter theta of the GMM by using the likelihood function of the characteristics of the maximum training set learned by the EM algorithm, wherein the parameter theta is expressed as
Figure BDA0001212440140000068
The EM algorithm is to calculate the solution by using two steps alternately, wherein the first step expectation (E) is used for calculating the expectation value of the log likelihood, and the second step maximization (M) is used for updating the parameters to find the parameter value of the maximization likelihood expectation, which is as follows:
e, step E:
Figure BDA0001212440140000069
Figure BDA00012124401400000610
Figure BDA00012124401400000611
wherein the content of the first and second substances,
Figure BDA00012124401400000612
is the kth Gaussian component generation feature fiA posterior probability of (d);
and M: updating the parameter Θ
Figure BDA00012124401400000613
Figure BDA00012124401400000614
Figure BDA00012124401400000615
Example 4:
as shown in fig. 4, the specific method for screening the candidate object includes: the method comprises the steps of utilizing an offline-trained cascade classifier to detect pedestrians in a specific scene, setting a low detection threshold in order to obtain more detection objects, and ensuring that all pedestrian images are obtained, wherein a large number of false alarms exist, and all candidate objects in the specific scene are assumed to be T ═ T { (T) } TiI 1.. M }, each candidate outputting a corresponding confidence score S ═ S ·iI 1., M }, sorting the confidence scores in descending order, setting two thresholds λhAnd λlRepresenting positive and negative examples with high and low confidence scores, respectively, H and N are defined as follows:
H＝{ti:si＞λh,i＝1,...,M}
N＝{ti:si＜λl,i＝1,...,M}
we expect a balanced set of data H and N, so let C ═ min (| H |, | N |) then
H＝{t1,...,tC}
N＝{tM-C+1,...,tM}
C＜M/2
Example 5:
as shown in fig. 5, the specific method of the online classifier is as follows: aiming at positive and negative sample sets H and N, firstly, expressing the positive and negative samples again by using an off-line trained Gaussian mixture model, then, training a classifier on line by using a LibSVM, and finally, using the on-line classifier to perform on-line classification on a candidate object tiIs a positive probability estimate. The method comprises the following specific steps:
step 1) characterization: for a single pedestrian sample feature is expressed as
Figure BDA0001212440140000071
Performing feature representation by using a Gaussian mixture model, wherein each Gaussian component is derived from a feature vector
Figure BDA0001212440140000072
Select a corresponding feature
Figure BDA0001212440140000076
Figure BDA0001212440140000073
For a single pedestrian sample picture, K features are selected by K Gaussian components, and the single pedestrian sample feature is re-expressed as [ f [ ]g1,fg2,...,fgK]And only the HOG characteristic is reserved for the final pedestrian sample characteristic, the position information of the image block is removed, and the final pedestrian sample characteristic is expressed as [ HOGg1,hogg2,...,hoggK]；
Step 2) training an SVM classifier: the SVM classifier is represented as
Figure BDA0001212440140000074
Wherein the kernel function adopts a Gaussian RBF kernel with the formula of
Figure BDA0001212440140000075
Step 3) prediction of an SVM classifier: using the classifier trained in the step 2) to perform all the candidate objects T ═ TiAnd predicting the I1, the other words, M, and outputting a probability estimation result.

Claims (4)

1. The pedestrian detection method based on self-learning is characterized by comprising the following steps: the method comprises the following specific steps: firstly, training an AdaBoost-based cascade classifier as an offline classifier, simultaneously training a Gaussian mixture model by using a group of public pedestrian photos, adopting HOG (histogram of oriented gradient) features and position information for feature coding, then adopting a low-threshold offline classifier to detect pedestrians in a specific scene, outputting confidence scores of candidate objects, then selecting high confidence scores as positive samples, taking low confidence scores as negative samples, re-representing the candidate detection objects by using the Gaussian mixture model, finally online training a discriminative pedestrian classifier by using an SVM (support vector machine) classifier, re-predicting the candidate objects and outputting probability estimation;
the offline classifier in the above step: training data come from any pedestrian data set, and feature extraction adopts LBP features;
the Gaussian mixture model is as follows: training a Gaussian mixture model by adopting an INRIA data set in an off-line manner, learning parameters of the Gaussian mixture model by adopting an EM (effective electromagnetic modeling) algorithm, and for each pedestrian sample, extracting HOG (histogram of oriented gradient) features and position information of each image block on a multi-scale image to form a group of feature vectors of the pedestrian sample;
screening candidate objects: setting a low threshold value of an off-line classifier, performing pedestrian detection on a specific scene by using the off-line classifier to obtain candidate detection objects and confidence scores thereof, setting two threshold values to screen all candidate objects, and obtaining a positive and negative sample set;
an online classifier: for each pedestrian sample, regenerating a new feature representation by using a Gaussian mixture model, then training an SVM classifier on line, finally detecting candidate objects of a specific scene again by using the SVM classifier, and outputting probability estimation;
the Gaussian mixture model comprises two parts of feature coding and GMM training, and specifically comprises the following steps:
feature coding: aiming at each pedestrian picture in the INRIA data set, firstly constructing three layers of Gaussian pyramids, and then extracting overlapped image blocks from each layer of image pyramids; assume that each pedestrian picture comprises N image blocks
Figure FDA0002385335980000011
Extracting HOG feature HOG of each image blockpiAnd its location information lpi＝[x y]TFinally, the feature of each image block is coded as fpi＝[hogpi T,lpi T]TAll image blocks constitute pedestrian sample features
Figure FDA0002385335980000012
And GMM training: an off-line trained Gaussian mixture model is expressed as
Figure FDA0002385335980000013
Wherein K is the number of Gaussian mixture components,
Figure FDA0002385335980000014
i is a unit matrix of the image data,
Figure FDA0002385335980000015
is the mixing weight of the kth gaussian component,
Figure FDA0002385335980000016
is a Gaussian distribution with a mean value of μkAnd the variance is
Figure FDA0002385335980000017
f is the pedestrian sample feature in the feature encoding portion;
a set of pedestrian sample characteristics χ ═ f given in the INRIA dataset1,f2,...,fMAnd learning the parameter theta of the GMM by using the likelihood function of the characteristics of the maximum training set learned by the EM algorithm, wherein the parameter theta is expressed as
Figure FDA0002385335980000021
The EM algorithm is to calculate the solution by using two steps, wherein the first step expectation (E) is used for calculating the expectation value of the log likelihood, and the second step maximization (M) is used for updating the parameters to find the parameter value maximizing the likelihood expectation, which is as follows:
e, step E:
Figure FDA0002385335980000022
Figure FDA0002385335980000023
Figure FDA0002385335980000024
wherein the content of the first and second substances,
Figure FDA0002385335980000025
is the kth Gaussian component generation feature fiA posterior probability of (d);
and M: updating the parameter Θ
Figure FDA0002385335980000026
Figure FDA0002385335980000027
Figure FDA0002385335980000028
Where M is M-1, where M is the number of pedestrian samples in the INRIA dataset.
2. The self-learning based pedestrian detection method according to claim 1, wherein: the offline classifier is trained by adopting a cascade classifier in OpenCV, and comprises two parts of data set preparation and training program operation, wherein the data set preparation is training data: creating a positive sample set by using opencv _ createsamples, and manually preparing a negative sample picture; the training program is run to train the cascade classifier: setting the feature type as LBP, and training a cascade classifier by adopting opencv _ traincacade.
3. The self-learning based pedestrian detection method according to claim 1, wherein: the specific method for screening the candidate objects comprises the following steps: the method comprises the steps of utilizing an offline-trained cascade classifier to detect pedestrians in a specific scene, setting a low detection threshold in order to obtain more detection objects, and ensuring that all pedestrian images are obtained, wherein a large number of false alarms exist, and all candidate objects in the specific scene are assumed to be T ═ T { (T) } TiI 1.. M }, each candidate outputting a corresponding confidence score S ═ S ·iI 1., M }, sorting the confidence scores in descending order, setting two thresholds λhAnd λlRepresenting positive and negative examples with high and low confidence scores, respectively, H and N are defined as follows:
H＝{ti:si＞λh,i＝1,...,M}
N＝{ti:si＜λl,i＝1,...,M}
we expect a balanced set of data H and N, so let C ═ min (| H |, | N |) then
H＝{t1,...,tC}
N＝{tM-C+1,...,tM}
C＜M/2。
4. The self-learning based pedestrian detection method according to claim 3, characterized in that: the specific method of the online classifier comprises the following steps: aiming at positive and negative sample sets H and N, firstly, expressing the positive and negative samples again by using an off-line trained Gaussian mixture model, then, training a classifier on line by using a LibSVM, and finally, using the on-line classifier to perform on-line classification on a candidate object tiA probability estimate that is positive; the method comprises the following specific steps:
step 1) characterization: for a single pedestrian sample feature is expressed as
Figure FDA0002385335980000031
Performing feature representation by using a Gaussian mixture model, wherein each Gaussian component is derived from a feature vector
Figure FDA0002385335980000032
Select a corresponding feature
Figure FDA0002385335980000033
Figure FDA0002385335980000034
For a single pedestrian sample picture, K features are selected by K Gaussian components, and the single pedestrian sample feature is re-expressed as [ f [ ]g1,fg2,...,fgK]And only the HOG characteristic is reserved for the final pedestrian sample characteristic, the position information of the image block is removed, and the final pedestrian sample characteristic is expressed as [ HOGg1,hogg2,...,hoggK]；
Step 2) training an SVM classifier: the SVM classifier is represented as
Figure FDA0002385335980000035
Wherein the kernel function adopts a Gaussian RBF kernel with the formula of
Figure FDA0002385335980000036
Step 3) prediction of an SVM classifier: using the classifier trained in the step 2) to perform all the candidate objects T ═ TiAnd predicting the I1, the other words, M, and outputting a probability estimation result.
CN201710033677.9A 2017-01-18 2017-01-18 Pedestrian detection method based on self-learning Expired - Fee Related CN106845387B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710033677.9A CN106845387B (en) 2017-01-18 2017-01-18 Pedestrian detection method based on self-learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710033677.9A CN106845387B (en) 2017-01-18 2017-01-18 Pedestrian detection method based on self-learning

Publications (2)

Publication Number Publication Date
CN106845387A CN106845387A (en) 2017-06-13
CN106845387B true CN106845387B (en) 2020-04-24

Family

ID=59123725

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710033677.9A Expired - Fee Related CN106845387B (en) 2017-01-18 2017-01-18 Pedestrian detection method based on self-learning

Country Status (1)

Country Link
CN (1) CN106845387B (en)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108364192B (en) * 2018-01-16 2022-10-18 创新先进技术有限公司 User mining method and device and electronic equipment
CN108875770B (en) * 2018-02-06 2021-11-19 北京迈格威科技有限公司 Pedestrian detection false alarm data labeling method, device, system and storage medium
CN108629309A (en) * 2018-04-28 2018-10-09 成都睿码科技有限责任公司 Foundation pit surrounding people's method for protecting
CN108875574B (en) * 2018-05-11 2021-06-25 北京旷视科技有限公司 Method, device and system for detecting false alarm result of pedestrian detection and storage medium
CN108647644B (en) * 2018-05-11 2021-04-06 山东科技大学 Coal mine blasting unsafe action identification and judgment method based on GMM representation
CN109492522B (en) * 2018-09-17 2022-04-01 中国科学院自动化研究所 Specific object detection model training program, apparatus, and computer-readable storage medium
US11715111B2 (en) * 2018-09-25 2023-08-01 Capital One Services, Llc Machine learning-driven servicing interface
CN113963456A (en) * 2020-07-20 2022-01-21 核工业理化工程研究院 Method and system for analyzing operation data of multiple high-speed rotating devices

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102609682A (en) * 2012-01-13 2012-07-25 北京邮电大学 Feedback pedestrian detection method for region of interest
CN103593672A (en) * 2013-05-27 2014-02-19 深圳市智美达科技有限公司 Adaboost classifier on-line learning method and Adaboost classifier on-line learning system
CN104156734A (en) * 2014-08-19 2014-11-19 中国地质大学（武汉） Fully-autonomous on-line study method based on random fern classifier
CN104517127A (en) * 2013-09-27 2015-04-15 汉王科技股份有限公司 Self-learning pedestrian counting method and apparatus based on Bag-of-features model

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102609682A (en) * 2012-01-13 2012-07-25 北京邮电大学 Feedback pedestrian detection method for region of interest
CN103593672A (en) * 2013-05-27 2014-02-19 深圳市智美达科技有限公司 Adaboost classifier on-line learning method and Adaboost classifier on-line learning system
CN104517127A (en) * 2013-09-27 2015-04-15 汉王科技股份有限公司 Self-learning pedestrian counting method and apparatus based on Bag-of-features model
CN104156734A (en) * 2014-08-19 2014-11-19 中国地质大学（武汉） Fully-autonomous on-line study method based on random fern classifier

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于快速增量学习的行人检测方法;施培蓓;《小型微型计算机系统》;20150831;第36卷(第8期);第1837-1841页 *

Also Published As

Publication number Publication date
CN106845387A (en) 2017-06-13

Similar Documents

Publication Publication Date Title
CN106845387B (en) Pedestrian detection method based on self-learning
CN107563372B (en) License plate positioning method based on deep learning SSD frame
CN110619369B (en) Fine-grained image classification method based on feature pyramid and global average pooling
Bahnsen et al. Rain removal in traffic surveillance: Does it matter?
CN103761531B (en) The sparse coding license plate character recognition method of Shape-based interpolation contour feature
CN106778796B (en) Human body action recognition method and system based on hybrid cooperative training
CN109190479A (en) A kind of video sequence expression recognition method based on interacting depth study
CN110826429A (en) Scenic spot video-based method and system for automatically monitoring travel emergency
CN106257490A (en) The method and system of detection driving vehicle information
CN113361464A (en) Vehicle weight recognition method based on multi-granularity feature segmentation
CN110008899B (en) Method for extracting and classifying candidate targets of visible light remote sensing image
CN105893971A (en) Traffic signal lamp recognition method based on Gabor and sparse representation
CN111915583A (en) Vehicle and pedestrian detection method based on vehicle-mounted thermal infrared imager in complex scene
WO2024051296A1 (en) Method and apparatus for obstacle detection in complex weather
CN114049572A (en) Detection method for identifying small target
CN109543498B (en) Lane line detection method based on multitask network
Orozco et al. Vehicular detection and classification for intelligent transportation system: A deep learning approach using faster R-CNN model
Sathya et al. Perspective vehicle license plate transformation using deep neural network on genesis of CPNet
Hua et al. Traffic lights detection and recognition method using deep learning with improved YOLOv5 for autonomous vehicle in ROS2
Nakamura et al. Few-shot adaptive object detection with cross-domain cutmix
CN111832463A (en) Deep learning-based traffic sign detection method
CN112750128A (en) Image semantic segmentation method and device, terminal and readable storage medium
CN103164707A (en) Shot boundary detection method based on support vector machine and particle swarm optimization algorithm
CN115909276A (en) Improved YOLOv 5-based small traffic sign target detection method in complex weather
CN114898290A (en) Real-time detection method and system for marine ship

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
CF01 Termination of patent right due to non-payment of annual fee
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20200424