CN109409190A - Pedestrian detection method based on histogram of gradients and Canny edge detector - Google Patents
Pedestrian detection method based on histogram of gradients and Canny edge detector Download PDFInfo
- Publication number
- CN109409190A CN109409190A CN201810954623.0A CN201810954623A CN109409190A CN 109409190 A CN109409190 A CN 109409190A CN 201810954623 A CN201810954623 A CN 201810954623A CN 109409190 A CN109409190 A CN 109409190A
- Authority
- CN
- China
- Prior art keywords
- image
- pixel
- point
- sample
- gradient
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/13—Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30196—Human being; Person
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (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 present invention provides a kind of pedestrian detection method based on histogram of gradients and Canny edge detector, method once carry out the processing of image gray processing, gamma correction, image gaussian filtering to the image of acquisition comprising steps of step 1;Step 2, the marginal information of image is extracted using Canny edge detector, calculates the histogram of gradients information in image, obtains the feature vector of description pedestrian's feature;Step 3, the positive sample comprising human head and shoulder part and the negative sample Training Support Vector Machines classifier not comprising human head and shoulder part are utilized;Step 4, the local window of all images is detected using the classifier that training obtains and is judged in window with the presence or absence of human head and shoulder target.
Description
Technical field
The present invention relates to a kind of image processing techniques, especially a kind of to be based on histogram of gradients and Canny edge detector
Pedestrian detection method.
Background technique
With the rapid development of artificial intelligence, machine vision is in an increasingly wide range of applications in real life.And people
Automatic search and rescue etc. of the physical examination method of determining and calculating under automatic Pilot, intelligent monitoring and disaster environment suffer from good application prospect.People
Class life environment it is complicated and changeable, human body also be not changeless rigid body, on the contrary, human body flexibly construct be human body form
It is very various.The diversification of the complicated and changeable and human figure of background environment makes human testing difficulty with higher.
The pedestrian detection method of mainstream is the characteristic information by extracting human body at present, then with the training of these characteristic informations
Classifier passes through the human body target in trained detection of classifier image.Representative algorithm has HOG (histogram of gradients) special
Sign, LBP (local binary patterns) feature etc..This category feature is mainly according to the profile information of human body, and texture information etc. describes
Body shape.When human body is there are when partial occlusion or larger human body deviation upright state, detection effect is substantially reduced.Classifier
SVM (support vector machines) mainly is used, the methods of Adboost.The development of algorithm deep learning, neural network algorithm.But it calculates
Method complexity is very high, and operand is huge.
Image border is calculated using discrete first derivative in traditional histogram of gradients pedestrian detection algorithm, the edge of extraction compared with
To be coarse, the inapparent position of change of gradient, the edge of extraction is discontinuous, includes more noise in edge image, is unfavorable for pair
The accurate description of pedestrian contour in image influences subsequent pedestrian detection effect.Traditional histogram of gradients pedestrian detection algorithm
Using the detection method of whole body, vertical features information accounts for very big specific gravity in pedestrian contour feature, this makes in detection process more
It is easy existing columnar object erroneous detections a large amount of in environment to be pedestrian.
Summary of the invention
The purpose of the present invention is to provide a kind of pedestrian detection side based on histogram of gradients and Canny edge detector
Method, the present invention can accurately describe characteristics of human body's information, and using detection of classifier pedestrian target and avoid believing due to human body
Cease missing inspection caused by partial occlusion.
Realize the technical solution of the object of the invention are as follows: a kind of pedestrian based on histogram of gradients and Canny edge detector
Detection method, which is characterized in that method comprising steps of
Step 1, the processing of image gray processing, gamma correction, image gaussian filtering is once carried out to the image of acquisition;
Step 2, the marginal information of image is extracted using Canny edge detector, calculates the histogram of gradients letter in image
Breath obtains the feature vector of description pedestrian's feature;
Step 3, the positive sample comprising human head and shoulder part and the training branch of the negative sample not comprising human head and shoulder part are utilized
Hold vector machine classifier;
Step 4, the local window of all images is detected using the obtained classifier of training and judge be in window
It is no that there are human head and shoulder targets.
Compared with prior art, the present invention having the advantage that (1) method proposed by the present invention is examined using the edge Canny
Device is surveyed, during calculating edge image, which is able to detect that finer image border, and marginal information is completeer
Whole, Gaussian filter reduces the interference of high-frequency noise, thus pedestrian contour is described it is more acurrate, to improve pedestrian detection
Effect;(2) method proposed by the present invention uses the local detection method of human head and shoulder part, only detects pedestrian's head-and-shoulder area,
The special chamfered shape of human head and shoulder part is more representative, can reduce the probability of erroneous detection;Human head and shoulder part can
The shape to keep relative stability, even if pedestrian is not for upright state or there are partial occlusions, it is also possible to detect row
People.Therefore method proposed by the present invention can be improved verification and measurement ratio.
The invention will be further described with reference to the accompanying drawings of the specification.
Detailed description of the invention
The present invention is based on the flow charts of histogram of gradients and the pedestrian detection method of Canny edge detector for the position Fig. 1.
Fig. 2 (a) is original image;Fig. 2 (b) is the edge detection effect of the edge detector in traditional HOG algorithm;Fig. 2
(c) be Canny edge detector in the method for the present invention edge detection effect.
Fig. 3 is that part sample in established positive sample library is trained in the method for the present invention.
Fig. 4 is the training process of the method for the present invention.
Fig. 5 is the testing process of the method for the present invention.
Fig. 6 (a) is the correct testing result in part of the method for the present invention;Fig. 6 (b) is the part missing inspection knot of the method for the present invention
Fruit;Fig. 6 (c) is the part erroneous detection result of the method for the present invention.
Fig. 7 is the statistical result of traditional histogram of gradients algorithm and the method for the present invention pattern detection effect.
Specific embodiment
A kind of pedestrian detection method based on histogram of gradients and Canny edge detector, comprising the following steps:
Step 1, image is obtained, image is pre-processed, is gamma correction, image gray processing, image Gauss filter respectively
Wave, less illumination, influence of the noise to testing result.
Gray processing is carried out to image, image is transformed into gray space from RGB color space, is convenient for subsequent filtering processing,
Calculation amount is decreased simultaneously.
It is uneven to reduce image irradiation using gamma correction the grey scale mapping of image to the lesser range of brightness change
Influence to testing result, the mapping relations in image grayscale space are as follows:
fγ(x, y)=cfγ(x,y)
Wherein fγ(x, y) is the image after conversion, and f (x, y) indicates that gray level image, c and γ are constant, c=in the present invention
1, γ=0.5.
To the discrete first derivative of noise-sensitive involved in edge extracting, therefore high frequency is filtered out using gauss low frequency filter
Noise reduces the interference that image high-frequency noises extract true edge.Gaussian filtering is real by Gaussian function and image convolution
It is existing:
fg(x, y)=G (x, y) * fγ(x,y)
Wherein, fg(x, y) indicates that smoothed out image, * are convolution algorithm symbol, and G (x, y) is Gaussian function, and σ is Gauss
The standard deviation of function.
Step 2, the marginal information of image is extracted using Canny edge detector, calculates the histogram of gradients letter in image
Breath obtains the feature vector of description pedestrian's feature.
Adjacent using pixel each in image or so, neighbouring pixel calculates separately the pixel in the gradient in the direction x, y,
Obtain the gradient image g in the direction x, yx(x,y)、gy(x, y):
gx(x, y)=fg(x+1,y)-fg(x-1,y)
gy(x, y)=fg(x,y+1)-fg(x,y-1)
Wherein gx(x,y)、gy(x, y) is the gradient image in the direction x, y respectively.
Non- maximum value inhibits, and the gradient magnitude of point o, point o in gradient magnitude image m (x, y) taken out in image is
mo, corresponding gradient direction is α in gradient direction image α (x, y)o.It crosses o point and makees straight line, eight neighbours of the straight line and o point
The rectangle intersection that domain point is formed is in two sub-pix points of a, b.Line is clicked through using the eight neighborhood of the o point adjacent with a, b point to insert
Value, is calculated the value m of two sub-pix pointsaAnd mb.Correlation Centre point pixel moWith two sub-pix point maAnd mbIf mo> ma
And mo> mbIt sets up simultaneously, is then left edge pixel;Otherwise it is isolated point, gives up the pixel.To all pixels in gradient image
Point executes aforesaid operations, the image g after just obtaining non-maximum suppressionm(x,y)。
Double Thresholding Segmentation, is arranged a high threshold, and a Low threshold is split the image after non-maximum suppression.If
Pixel value is greater than high threshold, retains pixel value, is otherwise set to zero, obtains high threshold gradient image;If pixel value is greater than Low threshold
But it is less than high threshold, retains pixel value, be otherwise set to zero, obtain Low threshold gradient image.It calculates as follows:
Wherein gh(x, y) is high threshold image, gl(x, y) is Low threshold image, ThFor high threshold, TlFor Low threshold, high threshold
Value is three times of Low threshold.
Image border connection, high threshold image ghIn non-zero pixels point be determining marginal point, be defined as strong pixel,
Low threshold image glIn non-zero pixels point be uncertain marginal point, be defined as weak pixel.Take one in high threshold image
A strong pixel judges that the eight neighborhood of the pixel of Low threshold image corresponding with the position whether there is weak pixel,
If it exists, then existing weak pixel is set to strong pixel, then repeatedly aforesaid operations, is otherwise set to zero.Then in high threshold
Next strong pixel is taken in value image, repeats aforesaid operations, until traversing entire image.Finally strong in Low threshold image
Pixel is added in high threshold image, obtains final edge gradient image g (x, y), and direction gradient figure, which also correspondingly retains, to be corresponded to
Pixel, obtain gradient direction figure α (x, y).
The histogram of gradients of statistical picture calculates image feature vector.Steps are as follows:
First evenly divided for 9 directions of 0~180 ° of direction in space, the sliding window time of fixed size is then used
Gradient direction figure is gone through, the gradient direction of pixel in sliding window is counted on this 9 directions.Count on this 9 directions and be in order to
The whole trend for obtaining region inner margin, is quantified as feature vector for marginal information.
Cycling among windows are slided using the block of fixed size, each piece is divided into four units, carries out gradient to each unit
The statistics of histogram is voted.The value of ballot is corresponding gradient magnitude, and weight is determined by the position coordinates and gradient direction of pixel
It is fixed.By the comprehensive feature vector for obtaining one 36 dimension of the histogram of gradients counted in four units, when block has traversed entire window,
Must a higher dimension vector.The vector is normalized, feature vector has just been obtained.Repeat above step, window traversal
Entire image has just obtained the feature vector of all windows.Statistics refers to that histogram of gradients is got by statistical gradient direction
, the feature vector of this 36 dimension includes the statistical distribution of the information in this block inside gradient direction, that is, reflects the region inner margin
Whole trend.All pieces of combination of eigenvectors is a high dimension vector in one detection window.
The boundary position of each unit and adjacent unit also have as closing property, only calculate gradient using the pixel of active cell
Histogram and ignore its relationship between adjacent cells, will generating region aliasing effect, this can make the feature of subsequent calculations
Vector generates mutation.Therefore it needs with weight come correction result.Weight is arranged to each unit of block, to eliminate aliasing effect,
Method particularly includes:
Step S101, if the coordinate of a pixel is (x, y), the central pixel point of unit is where pixel (x, y)
(x1,y1), the gradient direction of pixel (x, y) is θ, and θ is located at θ1And θ2Between both direction, wherein θ1< θ2;
Step S102, the position weight difference that pixel (x, y) is voted to four units
Wherein, dxAnd dyThe respectively length and width of unit;
Step S103, θ1The ballot weight in direction isθ2The ballot weight in direction is
Step 3, using the positive sample comprising human head and shoulder part, the negative sample training not comprising human head and shoulder part is supported
Vector machine classifier.
Positive and negative sample image is made, the Positive training sample in cut data library only retains the head-and-shoulder area of human body, made
For positive sample.It to the negative training sample in database, is cut at random, obtained negative sample.
Positive negative sample is put on " 1 " and " 0 " label respectively, the feature vector of positive negative sample, combination are extracted according to step (3)
An eigenmatrix is formed, the label of positive negative sample correspondingly forms a corresponding matrix, two Input matrixes to support
Vector machine, training obtain a classifier, which is preliminary classifier.
Using preliminary detection of classifier negative sample training negative sample, the sample of classification error is considered as difficult sample, will
It is saved as difficult sample.
The label that difficult sample is put on to " 0 ", is added in negative sample, then utilizes support vector machines training classification again
Device obtains final classifier.
Step 4, the local window for the detection of classifier image that training obtains judges in window with the presence or absence of human head and shoulder mesh
Mark traverses whole image, obtains final result.
It due to the size of pedestrian in image and is not fixed, needs to carry out multi-scale transform to input picture, establishing has not
With the image pyramid of resolution ratio:
{i1,i2,i3,...,in}
Wherein inFor the image of different resolution.
To the image of each scale, the feature vector of image is calculated using step (2), then utilize step (3) trained
To classifier calculated and judged:
S=j1×m×km×1
Wherein j1×mFor the feature vector of the m dimension in window, km×1For the classifier of m dimension, T is detection threshold value.
The calculating for completing all windows in image, has just obtained Preliminary detection result.
Finally in order to avoid the repetition of the same target detects, box cluster is carried out to the result of Preliminary detection, that is, is passed through
The similarity of the target position detected, merging obtain single detection block to get final testing result has been arrived.
Below with reference to simulation example of the invention, the present invention is described further.
The present invention use Opencv3.3,2013 translation and compiling environment of Visual Studio, Windows764 bit manipulation system,
The CPU of Intel Core i3 2.4GHz, 4G memory experimental situation.Test sample is the test sample in INRIA sample database.The detection of this paper
Standard is that area be overlapped between the bounding box and the bounding box of label of testing result is that can be considered successfully to examine more than or equal to 50%
Measure target.
As shown in Figure 1, the first step carries out gamma correction to input picture, to reduce the variation pair of extraneous intensity of illumination
Edge detection process generates interference.Then the gradient image of image is calculated using Canny operator, gradient image includes gradient magnitude
Image and gradient direction image.Histograms of oriented gradients system is carried out to the region of each detection window size of gradient image later
Meter, the result of statistics are the contour feature information indicated in the window.For the ease of unified standard, need to direction gradient histogram
Figure normalization, obtains final feature vector.Finally by the classified calculating of SVM classifier obtain in each window whether someone
Testing result, comprehensive all windows as a result, just obtained the testing result of final entire image, and be marked.
Fig. 2 (a) is original image.Fig. 2 (b) is the edge detection effect of the edge detector in traditional HOG algorithm.Fig. 2
(c) be Canny edge detector in the method for the present invention edge detection effect.
Effect of the edge that Canny edge detector extracts as seen from Figure 2 than the edge detector in traditional HOG algorithm
Fruit is far better.There is less noise at the edge of Canny algorithm detection first, the edge detector detection in traditional HOG algorithm
To edge then have more noise.In addition the edge detection from the effect at human body contour outline edge, in traditional HOG algorithm
The testing result of device is in arm, shoulder, and compared with Multiple level, the lines at certain profiles have even disappeared for the lines of leg position
, and the edge lines of Canny algorithm are more complete, can clearly depict human body contour outline.
Fig. 3 is that part sample in established positive sample library is trained in the method for the present invention.The present invention is studied with French INRIA
Pedestrian detection database be raw data base, establish the human head and shoulder sample database of oneself.The sample database includes 1394
Positive sample and 1218 negative samples, the size of positive negative sample are 64 X 64, and wherein positive sample is cut by the training positive sample of INRIA
It cuts to obtain, negative sample is then that the training negative sample random shearing to INRIA obtains.
As shown in figure 4, the training process of classifier is as follows: the first step extracts feature vector, institute to positive sample and negative sample
Directed quantity forms a two-dimensional eigenvectors matrix, and the label of all samples forms an one-dimensional matrix, the mark of positive sample
To be indicated with " 1 ", the label of negative sample is indicated label with " 0 ".By the training of the two Input matrix support vector machines, one point is obtained
Class device removes detection negative sample with the classifier, target can be detected in negative sample, these extremely difficult samples of result detected
This, difficulty sample is added in training sample, trains again, just obtained final training parameter.
Fig. 5 illustrates the process of pedestrian detection.It first has to carry out multi-scale transform to picture, it is then real under different scale
It now detects, testing result necessarily has the repetition of the same target to detect, so final detection window is merged, obtains most
Whole testing result.
Fig. 6 illustrates the testing result of part sample.
Fig. 7 is the statistics of tradition HOG algorithm and the method for the present invention test effect.The method of the present invention uses
Test sample in INRIA database is tested, and test sample includes 288 pictures with someone, wherein
589 positive sample targets are marked altogether.For more preferable evaluation algorithms, we depict omission factor (1- recall rate or false negative/
(false negative+true positives)) and the erroneous detection number (FPPI) in every picture double logarithmic curve.It can be seen from the figure that of the invention
The accuracy rate of method be higher than traditional HOG algorithm.When FPPI is 1, the omission factor ratio HOG algorithm of the method for the present invention is low
6.1%.This is because the edge detection that the method for the present invention uses is accurate, and using part detection.Accurate edge improves
Detect the probability of pedestrian, part detection reduces the probability of erroneous detection.
Table 1 illustrates the runing time of the detection process of the method for the present invention and tradition HOG algorithm.With the picture of 640 X 480
As detection sample.5 groups of runing times and average time are listed in table, it can be seen that the method for the present invention and tradition HOG algorithm
It is time-consuming essentially identical, i.e. the method for the present invention complexity for not increasing detection.So meeting requirement of real-time.
Table 1
Claims (7)
1. a kind of pedestrian detection method based on histogram of gradients and Canny edge detector, which is characterized in that method includes step
It is rapid:
Step 1, the processing of image gray processing, gamma correction, image gaussian filtering is once carried out to the image of acquisition;
Step 2, the marginal information of image is extracted using Canny edge detector, is calculated the histogram of gradients information in image, is obtained
To the feature vector of description pedestrian's feature;
Step 3, using the positive sample comprising human head and shoulder part and the negative sample not comprising human head and shoulder part training support to
Amount machine classifier;
Step 4, the local window of all images is detected using the classifier that training obtains and judges whether deposit in window
In human head and shoulder target.
2. the method according to claim 1, wherein detailed process is as follows for step 1:
A, image image gray processing: is transformed into gray space from RGB color space;
B, gamma correction: by formula (1) by the grey scale mapping of image to the lesser range of brightness change
fγ(x, y)=cfγ(x,y) (1)
Wherein, fγ(x, y) is by the image after gamma correction, and f (x, y) is gray level image, and c and γ are constant and γ < c;
C, image gaussian filtering: realize that the convolution of Gaussian function and image obtains smooth image f by formula (2)g(x, y)=G
(x,y)*fγ(x,y) (2)
Wherein, fg(x, y) is smoothed image,
3. the method according to claim 1, wherein detailed process is as follows for step 2:
Step 2.1, the pixel is calculated separately in the gradient in the direction x, y using the adjacent pixel up and down of pixel each in image
gx(x,y)、gy(x,y)
gx(x, y)=fg(x+1,y)-fg(x-1,y) (3)
gy(x, y)=fg(x,y+1)-fg(x,y-1) (4)
Step 2.2, gradient magnitude image m (x, y) is calculated
Step 2.3, for the point o in gradient image, the rectangle frame that o point does that one forms with o point eight neighborhood point excessively intersects two
The straight line of point a, b calculate 3 points of o, a, b of gradient magnitude image mo、maAnd mb;
If mo> maAnd mo> mb, then retaining o point is edge pixel;Otherwise o point is isolated point, gives up the point;
Step 2.4, all pixels point total to gradient image executes step 2.3, the image g after obtaining non-maximum suppressionm(x,y);
Step 2.5, gmThe image border (x, y) is attached, and a high threshold T is arrangedhWith a Low threshold Tl, to image gm(x,
Y) it is split according to formula (6) (7)
Wherein, gh(x,y)、gl(x, y) is respectively high threshold image, Low threshold image;
Step 2.6, by ghMiddle non-zero pixels point is defined as strong pixel, glIf middle non-zero pixels point is defined as pixel;
To ghEach of strong pixel, judge Low threshold image g corresponding with the positionlPixel eight neighborhood
With the presence or absence of weak pixel;
If it exists, then existing weak pixel is set to strong pixel, is otherwise set to zero;
Strong pixel in Low threshold image is added in high threshold image, final edge gradient image g (x, y) is obtained,
Direction gradient figure also correspondingly retains corresponding pixel, obtains gradient direction figure α (x, y)
α (x, y)=tan-1(gy(x,y)/gx(x,y)) (8)
Step 2.7,0~180 ° of direction in space is evenly divided into 9 directions, ladder is traversed using the sliding window of fixed size
It spends directional diagram α (x, y), the gradient direction of pixel in sliding window is counted in this 9 directions;
For each sliding window, traversal is slided using the block of fixed size, the statistics of histogram of gradients is carried out to each unit
Ballot, wherein each piece is divided into four units, the value of ballot is corresponding gradient magnitude, weight by pixel position coordinates
It is determined with gradient direction;
By the comprehensive feature vector for obtaining one 36 dimension of the histogram of gradients counted in four units, when block has traversed entire window
Mouthful, obtain the vector of a higher dimension;
It normalizes the vector and obtains feature vector.
4. according to the method described in claim 3, weight is arranged to each unit of block it is characterized in that, in step 2.7, with
Aliasing effect is eliminated, method particularly includes:
Step 2.7.1, if the coordinate of a pixel is (x, y), the central pixel point of unit where pixel (x, y) is (x1,
y1), the gradient direction of pixel (x, y) is θ, and θ is located at θ1And θ2Between both direction, wherein θ1< θ2;
Step 2.7.2, the position weight difference that pixel (x, y) is voted to four units
Wherein, dxAnd dyThe respectively length and width of unit;
Step 2.7.3, θ1The ballot weight in direction isθ2The ballot weight in direction is
5. the method according to claim 3 or 4, which is characterized in that detailed process is as follows for step 3:
Step 3.1, to the Positive training sample in database, only retain the head-and-shoulder area of human body, as positive sample, and on table
Label 1;It to the negative training sample in database, is cut at random, obtained negative sample, and label 0 on table;
Step 3.2, the feature vector of positive negative sample is extracted according to step 2, combination forms an eigenmatrix, the mark of positive negative sample
Label correspondingly form a corresponding matrix, and two Input matrixes to support vector machines, training obtains a classifier, this point
Class device is preliminary classifier;
Step 3.3, using preliminary detection of classifier negative sample training negative sample, the sample of classification error is considered as difficult sample
This, is saved as difficult sample;
Step 3.4, the label that difficult sample is put on to 0, is added in negative sample, then utilizes support vector machines training point again
Class device obtains final classifier.
6. according to the method described in claim 5, it is characterized in that, step 4 detailed process is as follows:
Step 4.1, multi-scale transform is carried out to the image of input, establishes the image pyramid { i with different resolution1,i2,
i3,...,in, wherein inFor the image of different resolution;
Step 4.2, the feature vector of image is calculated the image of each scale of each window using step 2, then utilizes step
The classifier that rapid 3 training obtains is calculated and is judged
S=j1×m×km×1
Wherein, j1×mFor the feature vector of the m dimension in window, km×1For the classifier of m dimension, T is detection threshold value.
7. according to the method described in claim 6, it is characterized in that, the result obtained to 4.2 carries out box cluster, detailed process
Are as follows:
The resultful rectangle frame of institute is divided into several classes by step 4.2.1, and each classification is numbered with number;
Step 4.2.2 is considered as same class if the absolute value of the interpolation on two rectangle frames, four vertex is less than some value;
Step 4.2.3, sums and is averaged to the coordinate of all rectangle frames of each classification, and the corresponding rectangle frame of average value is
The even rectangular frame position of each classification;
Step 4.2.4 filters out the classification for being less than threshold value comprising rectangle frame quantity, retains and is more than or equal to threshold comprising rectangle frame quantity
The classification of value;
Step 4.2.5 filters out the small rectangle frame being nested in big rectangle frame, just obtains final result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810954623.0A CN109409190A (en) | 2018-08-21 | 2018-08-21 | Pedestrian detection method based on histogram of gradients and Canny edge detector |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810954623.0A CN109409190A (en) | 2018-08-21 | 2018-08-21 | Pedestrian detection method based on histogram of gradients and Canny edge detector |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109409190A true CN109409190A (en) | 2019-03-01 |
Family
ID=65464324
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810954623.0A Pending CN109409190A (en) | 2018-08-21 | 2018-08-21 | Pedestrian detection method based on histogram of gradients and Canny edge detector |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109409190A (en) |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110119720A (en) * | 2019-05-17 | 2019-08-13 | 南京邮电大学 | A kind of real-time blink detection and pupil of human center positioning method |
CN110705630A (en) * | 2019-09-27 | 2020-01-17 | 聚时科技(上海)有限公司 | Semi-supervised learning type target detection neural network training method, device and application |
CN110706245A (en) * | 2019-10-18 | 2020-01-17 | 合肥工业大学 | Image edge detection method of GM (1,1) prediction model based on accumulation method |
CN111415516A (en) * | 2020-03-30 | 2020-07-14 | 福建工程学院 | Vehicle exhaust monitoring method of global road network |
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 |
CN111829458A (en) * | 2020-07-20 | 2020-10-27 | 南京理工大学智能计算成像研究院有限公司 | Gamma nonlinear error correction method based on deep learning |
CN111898587A (en) * | 2020-08-14 | 2020-11-06 | 广州盈可视电子科技有限公司 | Video coding processing method and device |
CN111950566A (en) * | 2020-08-04 | 2020-11-17 | 国网安徽省电力有限公司电力科学研究院 | Rotation-invariant HOG infrared image power equipment identification method |
CN112061905A (en) * | 2020-09-04 | 2020-12-11 | 禾麦科技开发(深圳)有限公司 | Elevator auxiliary dispatching method and system |
CN112071006A (en) * | 2020-09-11 | 2020-12-11 | 湖北德强电子科技有限公司 | High-efficiency low-resolution image area intrusion recognition algorithm and device |
CN112484680A (en) * | 2020-12-02 | 2021-03-12 | 杭州中为光电技术有限公司 | Sapphire wafer positioning and tracking method based on circle detection |
CN112784828A (en) * | 2021-01-21 | 2021-05-11 | 珠海市杰理科技股份有限公司 | Image detection method and device based on direction gradient histogram and computer equipment |
CN112818853A (en) * | 2021-02-01 | 2021-05-18 | 中国第一汽车股份有限公司 | Traffic element identification method, device, equipment and storage medium |
CN113240706A (en) * | 2021-04-12 | 2021-08-10 | 湖北工业大学 | Intelligent tracking detection method for molten iron tailings in high-temperature environment |
CN113989313A (en) * | 2021-12-23 | 2022-01-28 | 武汉智博通科技有限公司 | Edge detection method and system based on image multidimensional analysis |
CN116740465A (en) * | 2023-07-07 | 2023-09-12 | 国医通(北京)科技发展有限公司 | Focus sorter and equipment based on peritoneal dialysis liquid image segmentation |
CN117173703A (en) * | 2023-11-02 | 2023-12-05 | 温州华嘉电器有限公司 | Isolating switch state identification method |
CN117315289A (en) * | 2023-11-28 | 2023-12-29 | 苏州翰微材料科技有限公司 | Aeroengine blade contour edge detection method based on image processing |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140198951A1 (en) * | 2013-01-17 | 2014-07-17 | Canon Kabushiki Kaisha | Image processing apparatus and image processing method |
CN103955673A (en) * | 2014-04-30 | 2014-07-30 | 南京理工大学 | Body recognizing method based on head and shoulder model |
CN103971135A (en) * | 2014-05-05 | 2014-08-06 | 中国民航大学 | Human body target detection method based on head and shoulder depth information features |
CN105261017A (en) * | 2015-10-14 | 2016-01-20 | 长春工业大学 | Method for extracting regions of interest of pedestrian by using image segmentation method on the basis of road restriction |
CN107169979A (en) * | 2017-05-11 | 2017-09-15 | 南宁市正祥科技有限公司 | A kind of method for detecting image edge of improvement Canny operators |
-
2018
- 2018-08-21 CN CN201810954623.0A patent/CN109409190A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140198951A1 (en) * | 2013-01-17 | 2014-07-17 | Canon Kabushiki Kaisha | Image processing apparatus and image processing method |
CN103955673A (en) * | 2014-04-30 | 2014-07-30 | 南京理工大学 | Body recognizing method based on head and shoulder model |
CN103971135A (en) * | 2014-05-05 | 2014-08-06 | 中国民航大学 | Human body target detection method based on head and shoulder depth information features |
CN105261017A (en) * | 2015-10-14 | 2016-01-20 | 长春工业大学 | Method for extracting regions of interest of pedestrian by using image segmentation method on the basis of road restriction |
CN107169979A (en) * | 2017-05-11 | 2017-09-15 | 南宁市正祥科技有限公司 | A kind of method for detecting image edge of improvement Canny operators |
Non-Patent Citations (1)
Title |
---|
冯莹莹,郭常山著: "《智能监控视频中运动目标跟踪方法研究》", 30 June 2018 * |
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110119720A (en) * | 2019-05-17 | 2019-08-13 | 南京邮电大学 | A kind of real-time blink detection and pupil of human center positioning method |
CN110705630A (en) * | 2019-09-27 | 2020-01-17 | 聚时科技(上海)有限公司 | Semi-supervised learning type target detection neural network training method, device and application |
CN110706245B (en) * | 2019-10-18 | 2021-09-21 | 合肥工业大学 | Image edge detection method of GM (1,1) prediction model based on accumulation method |
CN110706245A (en) * | 2019-10-18 | 2020-01-17 | 合肥工业大学 | Image edge detection method of GM (1,1) prediction model based on accumulation method |
CN111415516A (en) * | 2020-03-30 | 2020-07-14 | 福建工程学院 | Vehicle exhaust monitoring method of global road network |
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 |
CN111829458A (en) * | 2020-07-20 | 2020-10-27 | 南京理工大学智能计算成像研究院有限公司 | Gamma nonlinear error correction method based on deep learning |
CN111950566A (en) * | 2020-08-04 | 2020-11-17 | 国网安徽省电力有限公司电力科学研究院 | Rotation-invariant HOG infrared image power equipment identification method |
CN111898587A (en) * | 2020-08-14 | 2020-11-06 | 广州盈可视电子科技有限公司 | Video coding processing method and device |
CN112061905B (en) * | 2020-09-04 | 2022-08-05 | 禾麦科技开发(深圳)有限公司 | Elevator auxiliary dispatching method and system |
CN112061905A (en) * | 2020-09-04 | 2020-12-11 | 禾麦科技开发(深圳)有限公司 | Elevator auxiliary dispatching method and system |
CN112071006A (en) * | 2020-09-11 | 2020-12-11 | 湖北德强电子科技有限公司 | High-efficiency low-resolution image area intrusion recognition algorithm and device |
CN112484680B (en) * | 2020-12-02 | 2022-06-03 | 杭州中为光电技术有限公司 | Sapphire wafer positioning and tracking method based on circle detection |
CN112484680A (en) * | 2020-12-02 | 2021-03-12 | 杭州中为光电技术有限公司 | Sapphire wafer positioning and tracking method based on circle detection |
CN112784828A (en) * | 2021-01-21 | 2021-05-11 | 珠海市杰理科技股份有限公司 | Image detection method and device based on direction gradient histogram and computer equipment |
CN112784828B (en) * | 2021-01-21 | 2022-05-17 | 珠海市杰理科技股份有限公司 | Image detection method and device based on direction gradient histogram and computer equipment |
CN112818853A (en) * | 2021-02-01 | 2021-05-18 | 中国第一汽车股份有限公司 | Traffic element identification method, device, equipment and storage medium |
CN113240706A (en) * | 2021-04-12 | 2021-08-10 | 湖北工业大学 | Intelligent tracking detection method for molten iron tailings in high-temperature environment |
CN113989313A (en) * | 2021-12-23 | 2022-01-28 | 武汉智博通科技有限公司 | Edge detection method and system based on image multidimensional analysis |
CN116740465A (en) * | 2023-07-07 | 2023-09-12 | 国医通(北京)科技发展有限公司 | Focus sorter and equipment based on peritoneal dialysis liquid image segmentation |
CN116740465B (en) * | 2023-07-07 | 2024-05-17 | 国医通(北京)科技发展有限公司 | Focus sorter and equipment based on peritoneal dialysis liquid image segmentation |
CN117173703A (en) * | 2023-11-02 | 2023-12-05 | 温州华嘉电器有限公司 | Isolating switch state identification method |
CN117173703B (en) * | 2023-11-02 | 2024-01-16 | 温州华嘉电器有限公司 | Isolating switch state identification method |
CN117315289A (en) * | 2023-11-28 | 2023-12-29 | 苏州翰微材料科技有限公司 | Aeroengine blade contour edge detection method based on image processing |
CN117315289B (en) * | 2023-11-28 | 2024-02-09 | 苏州翰微材料科技有限公司 | Aeroengine blade contour edge detection method based on image processing |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109409190A (en) | Pedestrian detection method based on histogram of gradients and Canny edge detector | |
WO2019104767A1 (en) | Fabric defect detection method based on deep convolutional neural network and visual saliency | |
CN106570486B (en) | Filtered target tracking is closed based on the nuclear phase of Fusion Features and Bayes's classification | |
CN103942577B (en) | Based on the personal identification method for establishing sample database and composite character certainly in video monitoring | |
CN104715238B (en) | A kind of pedestrian detection method based on multi-feature fusion | |
CN113592845A (en) | Defect detection method and device for battery coating and storage medium | |
CN109086718A (en) | Biopsy method, device, computer equipment and storage medium | |
CN109635875A (en) | A kind of end-to-end network interface detection method based on deep learning | |
Gao et al. | Building extraction from RGB VHR images using shifted shadow algorithm | |
CN106295124A (en) | Utilize the method that multiple image detecting technique comprehensively analyzes gene polyadenylation signal figure likelihood probability amount | |
CN106530271B (en) | A kind of infrared image conspicuousness detection method | |
CN105740945A (en) | People counting method based on video analysis | |
CN108573499A (en) | A kind of visual target tracking method based on dimension self-adaption and occlusion detection | |
CN109685045A (en) | A kind of Moving Targets Based on Video Streams tracking and system | |
CN108830856B (en) | GA automatic segmentation method based on time series SD-OCT retina image | |
CN106157308A (en) | Rectangular target object detecting method | |
CN107480607A (en) | A kind of method that standing Face datection positions in intelligent recording and broadcasting system | |
CN109461163A (en) | A kind of edge detection extraction algorithm for magnetic resonance standard water mould | |
CN108734200A (en) | Human body target visible detection method and device based on BING features | |
CN106056078B (en) | Crowd density estimation method based on multi-feature regression type ensemble learning | |
Wang et al. | The PAN and MS image fusion algorithm based on adaptive guided filtering and gradient information regulation | |
Gao et al. | An effective retinal blood vessel segmentation by using automatic random walks based on centerline extraction | |
CN108073940A (en) | A kind of method of 3D object instance object detections in unstructured moving grids | |
CN105335688B (en) | A kind of aircraft model recognition methods of view-based access control model image | |
CN108154513A (en) | Cell based on two photon imaging data detects automatically and dividing method |
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 |