CN109886086B - Pedestrian detection method based on HOG (histogram of oriented gradient) features and linear SVM (support vector machine) cascade classifier - Google Patents
Pedestrian detection method based on HOG (histogram of oriented gradient) features and linear SVM (support vector machine) cascade classifier Download PDFInfo
- Publication number
- CN109886086B CN109886086B CN201910006652.9A CN201910006652A CN109886086B CN 109886086 B CN109886086 B CN 109886086B CN 201910006652 A CN201910006652 A CN 201910006652A CN 109886086 B CN109886086 B CN 109886086B
- Authority
- CN
- China
- Prior art keywords
- sample
- svm
- pedestrian
- classifier
- data set
- 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.)
- Active
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a pedestrian detection method based on HOG characteristics and a linear SVM cascade classifier, which is mainly used for solving the safety problem of an intelligent vehicle in the driving process by improving the accuracy of implementing pedestrian detection of the intelligent vehicle in the driving process. In order to break the limitation that pedestrians must keep an upright posture when the HOG characteristics of the pedestrians are extracted by a traditional pedestrian detection algorithm based on HOG and SVM, the invention provides a method which divides the pedestrians into three different body actions of standing, squatting and stooping, trains the three body actions as a positive sample data set into corresponding weak classifiers, integrates the three obtained weak classifiers into a strong classifier, and uses the strong classifier with a cascade structure as an intelligent vehicle pedestrian detection model.
Description
Technical Field
The invention relates to a method for detecting pedestrians by using a cascade classifier based on HOG feature multi-classifier integration in an intelligent driving environment, which is mainly used for solving the safety problem of road driving in an unmanned driving environment and belongs to an intelligent driving technology and a computer vision technology.
Background
In recent years, although a large number of algorithms and improved technologies are emerging in the aspect of pedestrian detection, and tests are performed on public data sets such as INRIA and Caltech, so that the detection accuracy of pedestrian detection is gradually improved, due to some difficulties existing in pedestrian detection, such as influence on pedestrian detection caused by complicated situations such as change of body motion of a pedestrian and pedestrian shielding, the pedestrian detection still has great challenges, and meanwhile, the problems are urgently solved. There are two main methods for detecting pedestrians: firstly, based on the traditional algorithm, the characteristics need to be designed manually, and the design process is relatively complex; and secondly, based on a deep learning algorithm, a series of work needs to be perfected, such as network optimization, parameter adjustment and the like.
Disclosure of Invention
The purpose of the invention is as follows: in order to overcome the defects in the prior art, the invention provides a pedestrian detection method based on HOG characteristics and a linear SVM cascade classifier in an intelligent driving environment, which is used for safe, accurate and efficient pedestrian detection and pedestrian marking. The method comprises the steps of firstly collecting pedestrian data of a driving environment through an intelligent vehicle body sensor, then processing the data according to a cascade classifier trained based on HOG and linear SVM to detect pedestrians on a road, and prompting a vehicle driver or achieving the purpose of avoiding the pedestrians through intelligent control. The method includes the steps that three classifiers are designed and integrated according to three limb action characteristics of a pedestrian in a standing position, a squatting position and a bending position, the three limb action characteristics are respectively used as positive samples, when one limb action characteristic is used as a positive sample data set, the other two positive samples and a background image are used as negative samples, so that the three classifiers can be obtained, and finally the three classifiers are combined into a classifier with a cascade structure and written into an intelligent driving system. The intelligent driving system can detect and identify the pedestrian through the cascade classifier and then mark the pedestrian target, so that the accuracy of pedestrian detection of the intelligent vehicle in the driving process is ensured.
The technical scheme is as follows: in order to achieve the purpose, the invention adopts the technical scheme that:
the invention aims to improve the accuracy of pedestrian detection in the driving process of an intelligent vehicle, and is mainly characterized by the real-time property, high efficiency and accuracy of a pedestrian detection algorithm in an intelligent driving system. Firstly, a large number of sample sets of pedestrians on a driving road need to be collected, and the sample sets are divided into positive sample sets and negative sample sets. If the positive sample set is a standing pedestrian, the squat pedestrian, the stooped pedestrian and the background are negative sample sets; if the positive sample set is a squat pedestrian, standing pedestrians, stooped pedestrians and the background are negative sample sets; if the positive sample set is a stooped pedestrian, the squat pedestrian, the standing pedestrian and the background are negative sample sets. The HOG characteristics of the positive sample set and the HOG characteristics of the negative sample set are calculated respectively, the linear SVM is used for training respectively, three classifiers are obtained, the three classifiers are combined into a classifier with a cascade structure, and the Adaboost algorithm is used for setting weight for training again. The weights of the three classifiers are set according to the probability of the limb action of the pedestrian in the walking process, and the higher the probability of the certain limb action of the pedestrian is, the higher the weight setting is. For example, a standing motion is a normal body motion of a pedestrian during walking, and thus the weight set by a standing pedestrian for a classifier of a positive sample set is higher. And then writing the trained cascade classifier into an intelligent driving system, taking pedestrian data in the surrounding environment of the intelligent vehicle as input, and taking intelligent prompt or intelligent control of the intelligent vehicle as output, thereby achieving the purpose of avoiding pedestrians. Through the safety strategy of marking the pedestrian target after detecting and identifying the pedestrian layer by layer, the false detection rate of the pedestrian detection can be greatly reduced, and the safety of the intelligent vehicle in the driving process is improved.
A pedestrian detection method based on HOG features and a linear SVM cascade classifier comprises the following steps:
the method comprises the following steps: the method comprises the following steps of collecting pictures of pedestrians with three limb actions of standing, squatting and stooping to form a sample set, and dividing the sample set into three types:
sample data set T1: taking a pedestrian picture of standing limb actions as a positive sample, and taking a pedestrian picture of squatting and stooping limb actions as a negative sample;
sample data set T2: taking a picture of a pedestrian with squat limbs as a positive sample, and taking pictures of pedestrians with standing and bending limbs as a negative sample;
sample data set T3: taking the pedestrian pictures of the stooping limb actions as positive samples, and taking the pedestrian pictures of the standing and squatting limb actions as negative samples;
in the sample data set T1、T2、T3In (2), a label α is assigned to each positive sample, and a label β is assigned to each negative sample;
step two: extracting HOG characteristics of each sample;
step three: setting initial weight, and collecting sample data set T1、T2、T3Firstly training a SVM weak classifier S by using initial weight for positive and negative samples in the system1(ii) a Then, the weight is updated according to the learning error rate, and the sample data set T is sampled1、T2、T3The positive and negative samples in the SVM are retrained by the updated weight to obtain an SVM weak classifier S2(ii) a Then, the weight is updated according to the new learning error rate, and the sample data set T is sampled1、T2、T3The positive and negative samples in the SVM are retrained by the updated weight to obtain an SVM weak classifier S3(ii) a Repeating the steps until the number of the SVM weak classifiers reaches the specified number;
step four: carrying out weight training on a specified number of SVM weak classifiers based on the Adaboost idea, and integrating into a cascade strong classifier;
step five: using cascaded strong classifiers only on sample data set T1、T2、T3Identifying the negative sample, and adding the image identified by mistake into the negative sample difficult sample set; and optimally training the cascade strong classifier by using a negative sample difficulty case set.
Specifically, in the first step, in addition to acquiring the pedestrian picture, a background picture is acquired at the same time, and the background picture is taken as a sample data set T1、T2、T3A medium negative sample.
Specifically, in the second step, the HOG feature extraction method of the sample includes: firstly, carrying out gray level processing on a picture; then carrying out Gamma correction on the picture after the gray processing; and finally, carrying out HOG feature extraction on the picture after Gamma correction, firstly dividing the picture into a plurality of cell units, then collecting the direction histograms of the gradients or edges of all pixel points in the cell units, and then combining the direction histograms to form the HOG feature of the sample.
Specifically, in the third step, the training method of the SVM weak classifier includes: sample data set T1、T2、T3The HOG characteristics and the labels of the positive and negative samples in the SVM are input into the SVM for training,and storing the final training result as an XML file, wherein the content of the XML file comprises two arrays and a floating point number, the two arrays are respectively recorded as a Support Vector and an Alpha, and the floating point number is recorded as a Rho.
Specifically, in the fourth step, the cascade strong classifier is:
wherein: sk(x) Is the kth SVM weak classifier, alphakIs Sk(x) K is the total number of the weak classifiers of the SVM, sign (·) is a symbolic function, and x is a data set; e.g. of the typekIs the error rate.
By adopting the intelligent automobile, the pedestrian detection method under the intelligent driving environment comprises the following steps:
(1) in the running process of the intelligent vehicle, sensing the surrounding environment through a camera, a radar or a multi-sensor fusion technology, providing surrounding environment data and writing the surrounding environment data into an intelligent driving system;
(2) the intelligent driving system detects and identifies pedestrians and then marks pedestrian targets according to data imported by the sensor by using the cascade classifier, displays the pedestrian targets to a driver by using a graphical interface, or directly transmits a control instruction for operating the intelligent vehicle to the intelligent control layer through the intelligent driving system;
(3) and finally, the intelligent vehicle provides necessary safety prompts for the driver according to the specific position of the pedestrian displayed by the intelligent driving system, such as: whistling, alarming, etc., or intelligently controlled to avoid detected pedestrians, such as: lane changing, acceleration and deceleration, etc.
Has the advantages that: compared with the prior art, the pedestrian detection method based on the HOG characteristic and the linear SVM cascade classifier has the following advantages: 1. aiming at the limitation that the pedestrian has to keep an upright posture when the HOG characteristics of the pedestrian are extracted, the pedestrian posture is divided into three body motion characteristics of standing, squatting and stooping, and the missing rate of pedestrian detection can be obviously reduced by the method of the linear SVM cascade classifier designed according to the characteristics; 2. the missing detection of the pedestrian is one of the important hidden dangers of safety problems in the driving process of the intelligent vehicle, the accuracy of pedestrian detection is improved by improving a pedestrian detection algorithm in an intelligent driving system, and the safety of the intelligent vehicle in the driving process can be improved; 3. although the pedestrian detection algorithm based on deep learning is high in efficiency, the requirements of the algorithm on the training data volume and the performance of the vehicle-mounted processor are far higher than those of the traditional algorithm, and therefore the improvement on the traditional algorithm is beneficial to saving the manufacturing cost of the intelligent vehicle.
Drawings
FIG. 1 is a diagram of an implementation process of an HOG feature extraction algorithm;
FIG. 2 is a flow chart of a linear SVM solution;
FIG. 3 is a flow chart of a basic idea of the Adaboost algorithm;
fig. 4 is a flow chart of pedestrian detection in intelligent driving.
Detailed Description
The present invention will be further described with reference to the accompanying drawings.
The pedestrian detection technology in the intelligent driving technology is a technology that an intelligent driving system judges whether a pedestrian exists in a given image and video and gives a specific position of the pedestrian. The pedestrian detection algorithm is mainly completed by three steps of HOG feature extraction, SVM training and cascade classifier training. Compared with a pedestrian detection method based on deep learning, the pedestrian detection method based on the HOG features and the linear SVM cascade classifier is adopted, the problem that the recognition rate is reduced due to overlarge actions of the limbs of the pedestrian can be effectively solved, and the pedestrian detection method based on the HOG features and the linear SVM cascade classifier has the advantages that the model is simple, and the training is convenient and flexible.
The Histogram of Oriented Gradients (HOG) feature is a feature description value used for object detection in computer vision and image processing, and its essence is statistical information of gradients. The main idea is that in one image, because the gradient exists mainly at the edge, the appearance and shape of the local target can be well described by the gradient, so we can form the feature of the target by calculating and counting the gradient of the local area of the image. HOG has its own advantages over other characterization methods. For example, in a large space domain, it is difficult to maintain the geometric deformation and optical deformation of the image, but the HOG is operated on the local grid cells of the image, so that it can maintain good invariance to the geometric deformation and optical deformation of the image. However, the extraction of the HOG features is only suitable for rough spatial sampling, fine directional sampling, strong local optical normalization and the like, so when the HOG is used for describing features of pedestrians, the pedestrians must keep a substantially upright posture, and if the limbs of the pedestrians are too large, such as squat, bending down and the like, the effect of feature description is seriously affected.
A Support Vector Machine (SVM) is an optimal classification algorithm. The optimal classification is to say that the classification line is required to separate two classes without errors, the classification interval between the two classes is the largest, the classification line is popularized to a high-dimensional space, and the optimal classification line becomes an optimal classification surface. SVMs are trained using the classification interval concept, which relies on pre-processing of the data, so that the original pattern needs to be first expressed in a higher dimensional space, and then a non-linear mapping is found in this higher dimensional space, so that the original data belonging to the two classes, respectively, can be separated by a hyperplane. The goal of mapping data from the original space to the higher dimensional space is to find an optimal classification surface to maximize the classification interval. In addition, the support vector machine algorithm can adopt different kernel functions, the main kernel functions include a Linear kernel function (Linear) and a gaussian kernel function (RBF), a classifier trained by the SVM algorithm adopting the Linear kernel function is a Linear classifier, and a classifier trained by the SVM algorithm adopting the gaussian kernel function is a nonlinear classifier. The linear classifier has simple prediction function and high classification speed, while the nonlinear classifier can support a great number of vectors in a high-dimensional space, and the classification speed is far lower than that of the linear classifier. Because the dimensionality of the pedestrian features extracted by the HOG is high, selecting a linear classifier is a good choice.
The Cascade (Cascade) structure classifier is a high-efficiency classifier, and is a hierarchical classifier which is formed by combining a plurality of simple weak classifiers according to a certain sequence and is used for replacing a large and complex integral classifier. A small part of key features are selected from a huge feature set by applying an Adaboost algorithm on the structure of a cascade classifier, a weak classifier is trained by setting weights, and then a strong classifier is combined according to the weights of the weak classifier, so that the classification effect of improving the detection rate is finally achieved.
The method comprises the steps of training a cascade classifier based on HOG characteristics and a linear SVM, firstly, calculating HOG description values of positive and negative sample images to form a characteristic vector matrix, wherein each characteristic vector has a class label vector which can indicate the category of the characteristic vector; then training by using a linear SVM method; the trained three linear SVM classifiers are stored as XML files, and weight training is carried out on the three classifiers by using an Adaboost algorithm to form a well-arranged cascade classifier which can be used for pedestrian detection. And finally, combining a difficult case to train again, wherein the difficult case refers to all detected rectangular frames when the classifier of the first training is used for carrying out human detection on the original negative sample image, the areas of the rectangular frames are all false alarms, the false alarm rectangular frames are stored as pictures and added into the initial negative sample set, and the training and the reintegration of the linear SVM are carried out again.
Firstly, feature extraction
The sample dataset is mainly derived from INRIA. And acquiring a positive sample set, and extracting the HOG characteristic description value of the pedestrian in the positive sample set. The positive sample image uses 96 x 160 pixel pictures of a pedestrian standing, squatting, or stooping. When the human body image segmentation device is used, 16 pixels are removed from the upper, the lower, the left and the right of the picture, and the human body image part with 64-128 pixels in the middle is cut. And acquiring a negative sample set, extracting HOG characteristic description values of the negative sample set, wherein the negative sample image can be obtained by randomly cutting an image without a detection target and has the same size of 64 × 128 pixels. The number of negative samples is typically much larger than the number of positive samples. The positive and negative sample sets are divided into three categories:
sample data set T1: taking a pedestrian picture of standing limb actions as a positive sample, and taking a pedestrian picture of squatting and stooping limb actions and a background picture as negative samples;
sample data set T2: taking a pedestrian picture of the squat limb movement as a positive sample, and taking a pedestrian picture of the standing and stooping limb movement and a background picture as negative samples;
sample data set T3: taking a pedestrian picture of bending limbs as a positive sample, and taking a pedestrian picture of standing and squatting limbs and a background picture as negative samples;
in the sample data set T1、T2、T3In (1), a label 1 is assigned to each positive swatch, and a label 0 is assigned to each negative swatch;
the HOG feature extraction of an image is divided into two steps, as shown in fig. 1:
the method comprises the following steps: image preprocessing, namely performing color space normalization processing on the collected images to achieve the noise reduction effect; comprises two steps:
(1) image graying, that is, for color pictures, RGB components are converted into grayscale images, and the method used is a weighted average method. Firstly, endowing three components of an image with different weights, then carrying out weighted average to make the values be gray values, wherein the expression is as follows:
f(i,j)＝VR.R(i,j)+VG.G(i.j)+VB.B(i,j)
the final graying formula can be obtained according to the weight set as follows:
Gray＝0.3×R+0.59×G+0.11×B
(2) when the image brightness is not uniform, the Gamma correction can increase or decrease the overall brightness. The method used for Gamma correction is the square root method, and the formula of the square root method is as follows:
Y(x,y)＝I(x,y)γ
step two: HOG feature extraction
The image is first divided into small connected regions, called cell units. And then acquiring the direction histogram of the gradient or edge of each pixel point in the cell unit. Finally, these histograms are combined to form the feature description value. The method comprises the following specific steps:
(1) the gradients of the horizontal coordinate and the vertical coordinate of the image are calculated, and the gradient direction value of each pixel position is calculated according to the gradients.
(2) A coding method is provided for local image regions that maintain a weak sensitivity to the pose and appearance of human objects in the image.
(3) Due to the variation of local illumination and the variation of foreground-background contrast, the variation range of gradient intensity is very large, which requires normalization of gradient intensity. Normalization can further compress lighting, shadows, and edges.
(4) The last step is to collect the HOG features from all the overlapping parts of the detection window and combine them into a final feature vector for training the classifier.
Two, SVM training
Sample data set T1、T2、T3Firstly training a SVM weak classifier S by using initial weight for positive and negative samples in the system1. The training method of the SVM weak classifier comprises the following steps: sample data set T1、T2、T3The HOG characteristics and the labels of the positive and negative samples in the SVM are input into the SVM for training, and a principle flow chart of the SVM is shown in FIG. 2; the final training result will be saved as an XML file, the content of which includes two arrays and a floating point number.
Then, the weight is updated according to the learning error rate, so that the SVM weak classifier S1The weight of the training sample with high learning error rate is higher, so that the samples with high error rate can be more emphasized in the following weak classifiers; sample data set T1、T2、T3The positive and negative samples in the SVM are retrained by the updated weight to obtain an SVM weak classifier S2(ii) a Then, the weight is updated according to the new learning error rate, and the sample data set T is sampled1、T2、T3The positive and negative samples in the SVM are retrained by the updated weight to obtain an SVM weak classifier S3。
Three, cascade classifier training
The three weak classifiers generated by the above steps are integrated by using an Adaboost algorithm, and a basic idea flow chart is shown in FIG. 3. And finally, integrating the 3 weak classifiers through a set strategy to obtain a final strong learner. The Adaboost set strategy adopts a weighted average method, and the final strong classifier is as follows:
wherein: sk(x) Is the kth SVM weak classifier, alphakIs Sk(x) K is the total number of the SVM weak classifiers, sign (·) is a symbolic function, and x is a sample set; e.g. of the typekIs the error rate.
Fourth, combine difficult retraining
The bootstrap algorithm is used for optimizing the classifier of the cascade structure, and the basic idea of the bootstrap method is that an initial negative sample set is used for training a model, then negative samples which are wrongly classified by the initial model are collected to form a negative sample difficult case set, and a new model is trained by using the negative sample difficult case set.
The optimized cascade classifier can be applied to a pedestrian detection module in intelligent driving, and the working flow of the pedestrian detection module comprising the cascade classifier is shown in fig. 4.
The above description is only of the preferred embodiments of the present invention, and it should be noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the invention and these are intended to be within the scope of the invention.
Claims (5)
1. A pedestrian detection method based on HOG features and a linear SVM cascade classifier is characterized in that: the method comprises the following steps:
the method comprises the following steps: the method comprises the following steps of collecting pictures of pedestrians with three limb actions of standing, squatting and stooping to form a sample set, and dividing the sample set into three types:
sample data set T1: taking a pedestrian picture of standing limb actions as a positive sample, and taking a pedestrian picture of squatting and stooping limb actions as a negative sample;
sample data set T2: taking a picture of a pedestrian with squat limbs as a positive sample, and taking pictures of pedestrians with standing and bending limbs as a negative sample;
sample data set T3: taking the pedestrian pictures of the stooping limb actions as positive samples, and taking the pedestrian pictures of the standing and squatting limb actions as negative samples;
in the sample data set T1、T2、T3In (2), a label α is assigned to each positive sample, and a label β is assigned to each negative sample;
step two: extracting HOG characteristics of each sample;
step three: setting initial weight, and collecting sample data set T1、T2、T3Firstly training a SVM weak classifier S by using initial weight for positive and negative samples in the system1(ii) a Then, the weight is updated according to the learning error rate, and the sample data set T is sampled1、T2、T3The positive and negative samples in the SVM are retrained by the updated weight to obtain an SVM weak classifier S2(ii) a Then, the weight is updated according to the new learning error rate, and the sample data set T is sampled1、T2、T3The positive and negative samples in the SVM are retrained by the updated weight to obtain an SVM weak classifier S3(ii) a Repeating the steps until the number of the SVM weak classifiers reaches the specified number;
step four: updating the weights of SVM weak classifiers with the appointed number based on the Adaboost idea, and integrating into a cascade strong classifier;
step five: using cascaded strong classifiers only on sample data set T1、T2、T3Negative inThe method comprises the steps of identifying, adding an image which is identified by mistake into a negative sample difficult case set; and optimally training the cascade strong classifier by using a negative sample difficulty case set.
2. The pedestrian detection method based on the HOG feature and the linear SVM cascade classifier according to claim 1, wherein: in the first step, except for collecting the pedestrian picture, a background picture is collected at the same time, and the background picture is used as a sample data set T1、T2、T3A medium negative sample.
3. The pedestrian detection method based on the HOG feature and the linear SVM cascade classifier according to claim 1, wherein: in the second step, the HOG feature extraction method of the sample comprises the following steps: firstly, carrying out gray level processing on a picture; then carrying out Gamma correction on the picture after the gray processing; and finally, carrying out HOG feature extraction on the picture after Gamma correction, firstly dividing the picture into a plurality of cell units, then collecting the direction histograms of the gradients or edges of all pixel points in the cell units, and then combining the direction histograms to form the HOG feature of the sample.
4. The pedestrian detection method based on the HOG feature and the linear SVM cascade classifier according to claim 1, wherein: in the third step, the training method of the SVM weak classifier comprises the following steps: sample data set T1、T2、T3The HOG characteristics and the labels of the positive and negative samples in the XML file are input into the SVM for training, and the final training result is stored as an XML file, wherein the content of the XML file comprises two arrays and a floating point number.
5. The pedestrian detection method based on the HOG feature and the linear SVM cascade classifier according to claim 1, wherein: in the fourth step, the cascade strong classifier is as follows:
wherein: sk(x) Is the kth SVM weak classifier, alphakIs Sk(x) K is the total number of the weak classifiers of the SVM, sign (·) is a symbolic function, and x is a data set; e.g. of the typekTo learn an error rate.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910006652.9A CN109886086B (en) | 2019-01-04 | 2019-01-04 | Pedestrian detection method based on HOG (histogram of oriented gradient) features and linear SVM (support vector machine) cascade classifier |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910006652.9A CN109886086B (en) | 2019-01-04 | 2019-01-04 | Pedestrian detection method based on HOG (histogram of oriented gradient) features and linear SVM (support vector machine) cascade classifier |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109886086A CN109886086A (en) | 2019-06-14 |
| CN109886086B true CN109886086B (en) | 2020-12-04 |
Family
ID=66925514
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910006652.9A Active CN109886086B (en) | 2019-01-04 | 2019-01-04 | Pedestrian detection method based on HOG (histogram of oriented gradient) features and linear SVM (support vector machine) cascade classifier |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109886086B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110516523B (en) * | 2019-06-18 | 2023-08-25 | 中国矿业大学 | Pedestrian detection method for dangerous area in mobile video inspection |
| CN111080524A (en) * | 2019-12-19 | 2020-04-28 | 吉林农业大学 | Plant disease and insect pest identification method based on deep learning |
| CN111060904B (en) * | 2019-12-25 | 2022-03-15 | 中国汽车技术研究中心有限公司 | Blind area monitoring method based on millimeter wave and vision fusion perception |
| CN111783604A (en) * | 2020-06-24 | 2020-10-16 | 中国第一汽车股份有限公司 | Vehicle control method, device and equipment based on target identification and vehicle |
| CN114463813B (en) * | 2022-01-21 | 2024-02-13 | 苏州浪潮智能科技有限公司 | HOG feature-based expression recognition method, system and related device |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2008039514A1 (en) * | 2006-09-26 | 2008-04-03 | Cytrx Corporation | Pharmaceutical compositions and methods for treating diseases associated with neurodegeneration |
| CN102663409A (en) * | 2012-02-28 | 2012-09-12 | 西安电子科技大学 | Pedestrian tracking method based on HOG-LBP |
| CN103211599A (en) * | 2013-05-13 | 2013-07-24 | 桂林电子科技大学 | Method and device for monitoring tumble |
| CN104318263A (en) * | 2014-09-24 | 2015-01-28 | 南京邮电大学 | Real-time high-precision people stream counting method |
| CN104463232A (en) * | 2014-12-30 | 2015-03-25 | 中山大学 | Density crowd counting method based on HOG characteristic and color histogram characteristic |
| CN104484680A (en) * | 2014-09-26 | 2015-04-01 | 徐晓晖 | Multi-model multi-threshold combined pedestrian detection method |
| CN104715238A (en) * | 2015-03-11 | 2015-06-17 | 南京邮电大学 | Pedestrian detection method based on multi-feature fusion |
| CN106339684A (en) * | 2016-08-26 | 2017-01-18 | 北汽福田汽车股份有限公司 | Pedestrian detection method, device and vehicle |
| CN106934380A (en) * | 2017-03-19 | 2017-07-07 | 北京工业大学 | A kind of indoor pedestrian detection and tracking based on HOG and MeanShift algorithms |
| CN106971158A (en) * | 2017-03-23 | 2017-07-21 | 南京邮电大学 | A kind of pedestrian detection method based on CoLBP symbiosis feature Yu GSS features |
| CN107122726A (en) * | 2017-04-19 | 2017-09-01 | 高新兴科技集团股份有限公司 | A kind of multi-pose pedestrian detection method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170213080A1 (en) * | 2015-11-19 | 2017-07-27 | Intelli-Vision | Methods and systems for automatically and accurately detecting human bodies in videos and/or images |
-
2019
- 2019-01-04 CN CN201910006652.9A patent/CN109886086B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2008039514A1 (en) * | 2006-09-26 | 2008-04-03 | Cytrx Corporation | Pharmaceutical compositions and methods for treating diseases associated with neurodegeneration |
| CN102663409A (en) * | 2012-02-28 | 2012-09-12 | 西安电子科技大学 | Pedestrian tracking method based on HOG-LBP |
| CN103211599A (en) * | 2013-05-13 | 2013-07-24 | 桂林电子科技大学 | Method and device for monitoring tumble |
| CN104318263A (en) * | 2014-09-24 | 2015-01-28 | 南京邮电大学 | Real-time high-precision people stream counting method |
| CN104484680A (en) * | 2014-09-26 | 2015-04-01 | 徐晓晖 | Multi-model multi-threshold combined pedestrian detection method |
| CN104463232A (en) * | 2014-12-30 | 2015-03-25 | 中山大学 | Density crowd counting method based on HOG characteristic and color histogram characteristic |
| CN104715238A (en) * | 2015-03-11 | 2015-06-17 | 南京邮电大学 | Pedestrian detection method based on multi-feature fusion |
| CN106339684A (en) * | 2016-08-26 | 2017-01-18 | 北汽福田汽车股份有限公司 | Pedestrian detection method, device and vehicle |
| CN106934380A (en) * | 2017-03-19 | 2017-07-07 | 北京工业大学 | A kind of indoor pedestrian detection and tracking based on HOG and MeanShift algorithms |
| CN106971158A (en) * | 2017-03-23 | 2017-07-21 | 南京邮电大学 | A kind of pedestrian detection method based on CoLBP symbiosis feature Yu GSS features |
| CN107122726A (en) * | 2017-04-19 | 2017-09-01 | 高新兴科技集团股份有限公司 | A kind of multi-pose pedestrian detection method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109886086A (en) | 2019-06-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109886086B (en) | Pedestrian detection method based on HOG (histogram of oriented gradient) features and linear SVM (support vector machine) cascade classifier | |
| CN107563372B (en) | License plate positioning method based on deep learning SSD frame | |
| CN108304873B (en) | Target detection method and system based on high-resolution optical satellite remote sensing image | |
| TWI651662B (en) | Image annotation method, electronic device and non-transitory computer readable storage medium | |
| CN110929593B (en) | Real-time significance pedestrian detection method based on detail discrimination | |
| CN109460704B (en) | Fatigue detection method and system based on deep learning and computer equipment | |
| CN106650731B (en) | Robust license plate and vehicle logo recognition method | |
| US7724962B2 (en) | Context adaptive approach in vehicle detection under various visibility conditions | |
| CN112464883A (en) | Automatic detection and identification method and system for ship target in natural scene | |
| US10445602B2 (en) | Apparatus and method for recognizing traffic signs | |
| JP7206082B2 (en) | Systems and methods for recognizing traffic signs | |
| CN110163069B (en) | Lane line detection method for driving assistance | |
| CN107480585B (en) | Target detection method based on DPM algorithm | |
| CN110659550A (en) | Traffic sign recognition method, traffic sign recognition device, computer equipment and storage medium | |
| CN111062292B (en) | Fatigue driving detection device and method | |
| CN110569782A (en) | Target detection method based on deep learning | |
| CN108918532A (en) | A kind of through street traffic sign breakage detection system and its detection method | |
| CN112232371A (en) | American license plate recognition method based on YOLOv3 and text recognition | |
| CN116188999B (en) | Small target detection method based on visible light and infrared image data fusion | |
| CN113963333B (en) | Traffic sign board detection method based on improved YOLOF model | |
| TWI696958B (en) | Image adaptive feature extraction method and its application | |
| CN105354547A (en) | Pedestrian detection method in combination of texture and color features | |
| CN114419603A (en) | Automatic driving vehicle control method and system and automatic driving vehicle | |
| CN112347967B (en) | Pedestrian detection method fusing motion information in complex scene | |
| CN117475353A (en) | Video-based abnormal smoke identification method and system |
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 | ||
| CB02 | Change of applicant information | ||
| CB02 | Change of applicant information |
Address after: 210000, 66 new model street, Gulou District, Jiangsu, Nanjing Applicant after: NANJING University OF POSTS AND TELECOMMUNICATIONS Address before: No. 30 Guangyue Road, Qixia Street, Qixia District, Nanjing City, Jiangsu Province, 210000 Applicant before: NANJING University OF POSTS AND TELECOMMUNICATIONS |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |