CN106778504A - A kind of pedestrian detection method - Google Patents

Abstract

The invention discloses a kind of pedestrian detection method, comprise the following steps：1）Source images are pre-processed by collection pedestrian's video image；2）The primary detection of pedestrian：Extract great amount of images and be used as training sample；Feature extraction is carried out with improved HOG feature extracting methods；A two-value grader is built according to characteristic is extracted, sample training is carried out to grader；It is identified by grader, obtains primary pedestrian detection result；3）Pedestrian's accurate detection：Extract great amount of images and be used as training sample；The extraction of color, gradient, histogram feature is carried out using Integration tunnel feature extracting method, adjacent scale feature is calculated using adjacent scale feature approximate calculation method；It is trained using Adaboost obtains final strong classifier for target classification；Primary pedestrian detection result images are classified using grader, accurate pedestrian detection result is obtained.

Description

A kind of pedestrian detection method

Technical field

The present invention relates to pedestrian detection technology field, and in particular to a kind of pedestrian detection method.

Background technology

The main body of pedestrian detection technology research is the main part of real world, therefore explores what is pursued as people Emphasis.With the development of information technology, pedestrian detection technology also being applied among reality slowly, such as intelligent monitor system, Vehicle-mounted accessory system, intelligent robot etc..Just because of the demand of social development, pedestrian detection technology field has attracted everybody's ginseng With the research to this field.It is pedestrian detection technology breakthrough that HOG features in 2005 are proposed for pedestrian detection.With 10 years afterwards derive numerous methods, all obtain great successes.But pedestrian detection has the difficult point of many in itself：

(1) pedestrian is used as non-rigid targets, the change that the aspect such as itself outward appearance, attitude all can be at any time, influence pedestrian's inspection The accuracy of survey；

(2) complicated background environment can cause pedestrian detection flase drop or missing inspection occur, how to remove complex background and disturb into It is the difficult point of research；

(3) pedestrian's highly dense under large scene, pedestrian blocks the complexity for causing pedestrian detection.

The content of the invention

The technical problem to be solved in the present invention is to provide a kind of pedestrian detection method.

A kind of pedestrian detection method, comprises the following steps：

S1：Source images are pre-processed by collection pedestrian's video image；

S2：The primary detection of pedestrian：

S2-1：Extract great amount of images and be used as training sample, positive sample is had to comprising pedestrian, and negative sample does not include pedestrian's The combination of various background pictures；

S2-2：Feature extraction is carried out with improved HOG feature extracting methods；

S2-3：A two-value grader is built according to characteristic is extracted, sample training is carried out to grader；

S2-4：Extract great amount of images and be used as detection sample, feature extraction is carried out with improved HOG feature extracting methods；

S2-5：It is identified by grader, obtains primary pedestrian detection result；

S3：Pedestrian's accurate detection：

S3-1：Extract great amount of images and be used as training sample, positive sample is had to comprising pedestrian, and negative sample does not include pedestrian's The combination of various background pictures；

S3-2：The extraction of color, gradient, histogram feature is carried out using Integration tunnel feature extracting method, using adjacent Scale feature approximate calculation method calculates adjacent scale feature；

S3-3：It is trained using Adaboost obtains final strong classifier for target classification；

S3-4：The primary pedestrian detection result images obtained to step S2, carry out Integration tunnel feature extracting method extraction Feature, the grader using step S3-3 is classified, and obtains accurate pedestrian detection result.

Further, the improved HOG feature extracting methods are specific as follows：

1) source images are carried out into greyscale transformation and obtains corresponding gray level image；

2) Gamma corrections i.e. color space normalized is carried out to the gray level image for obtaining；

3) gradient of each pixel is calculated, the profile information of target image is obtained；

4) N number of Cell (cell compartment) is divided an image into, then several adjacent Cs ell is combined into a Block (sides Block region), the histogram of gradients in each Cell is calculated, the Feature Descriptor connection that all of Cell in Block is obtained is just To the feature of each Block；

5) it is as follows histogram of gradients as the probability density function in comentropy definition：

The θ of formula_iThe ratio of the gradient number and total gradient number in a gradient direction interval is represented, ladder in Block is calculated Degree the Direction interval manhole ladder number of degrees and the ratio of total gradient number, obtain the information entropy of each block, special using comentropy as sample Take over the detection of the training and pedestrian target in grader for use.

Further, the gradient calculation is specific as follows：

Gradient in gray level image each pixel two-dimensional coordinate plane is calculated, formula is as follows：

Grad_x(x, y)=IM (x+1, y)-IM (x-1, y)

Grad_y(x, y)=IM (x, y+1)-IM (x, y-1),

Wherein, Grad_x、Grad_yWith the gradient that IM (x, y) is respectively abscissa, the gradient on ordinate, image intensity value Size, therefore, the gradient calculation of each pixel is as follows in image：

Wherein, Grad (x, y) is the gradient magnitude size of IM (x, y), and θ (x, y) is that IM (x, y) gradient direction angle is big It is small.

Further, the adjacent scale feature approximate calculation is as follows：

1) for source images Im (x, y) for giving, the factor of influence τ of picture up-sampling, by the image of up-sampling by Im' (x, y) is represented, the formula of conversion is as follows：

Im'(x, y)=Im (x/ τ, y/ τ)；

2) mode calculated according to image gradient can be seen that the gradient of image gradient information and source images after up-sampling There is following relation in information：

3) relation therefore between the total gradient magnitude of up-sampling image and the gradient magnitude of source images is as follows：

The beneficial effects of the invention are as follows：

1) present invention uses improved HOG features to be detected for pedestrian target in primary detection, in traditional HOG features Combining information entropy feature on the basis of detection, is favorably improved the accuracy rate of pedestrian detection.

2) present invention detects that Integration tunnel feature is main using Integration tunnel feature in accurate detection for pedestrian target Color, gradient magnitude and histogram of gradients information are merged, it is contemplated that feature is extracted in the case of Image Multiscale and spends more Time, it is proposed that the method that swift nature is calculated, the method for fast computing features is mainly the feature letter calculated under current scale Breath, then approximate calculation goes out the characteristic information of adjacent yardstick hypograph, in the situation of Preliminary detection exclusive segment non-pedestrian window Under, can more accurately detect pedestrian target.

Specific embodiment

The present invention is further elaborated for specific examples below, but not as a limitation of the invention.

A kind of pedestrian detection method, comprises the following steps：

S1：Source images are pre-processed by collection pedestrian's video image；

S2：The primary detection of pedestrian：

S2-1：Extract great amount of images and be used as training sample, positive sample is had to comprising pedestrian, and negative sample does not include pedestrian's The combination of various background pictures；

S2-2：Feature extraction is carried out with improved HOG feature extracting methods；

S2-3：A two-value grader is built according to characteristic is extracted, sample training is carried out to grader；

S2-4：Extract great amount of images and be used as detection sample, feature extraction is carried out with improved HOG feature extracting methods；

S2-5：It is identified by grader, obtains primary pedestrian detection result；

S3：Pedestrian's accurate detection：

S3-1：Extract great amount of images and be used as training sample, positive sample is had to comprising pedestrian, and negative sample does not include pedestrian's The combination of various background pictures；

S3-2：The extraction of color, gradient, histogram feature is carried out using Integration tunnel feature extracting method, using adjacent Scale feature approximate calculation method calculates adjacent scale feature；

S3-3：It is trained using Adaboost obtains final strong classifier for target classification；

S3-4：The primary pedestrian detection result images obtained to step S2, carry out Integration tunnel feature extracting method extraction Feature, the grader using step S3-3 is classified, and obtains accurate pedestrian detection result.

Improved HOG feature extracting methods are specific as follows：

1) source images are carried out into greyscale transformation and obtains corresponding gray level image；

2) Gamma corrections i.e. color space normalized is carried out to the gray level image for obtaining；

3) gradient of each pixel is calculated, the profile information of target image is obtained；

4) N number of Cell (cell compartment) is divided an image into, then several adjacent Cs ell is combined into a Block (sides Block region), the histogram of gradients in each Cell is calculated, the Feature Descriptor connection that all of Cell in Block is obtained is just To the feature of each Block；

5) it is as follows histogram of gradients as the probability density function in comentropy definition：

The θ of formula_iThe ratio of the gradient number and total gradient number in a gradient direction interval is represented, ladder in Block is calculated Degree the Direction interval manhole ladder number of degrees and the ratio of total gradient number, obtain the information entropy of each block, special using comentropy as sample Take over the detection of the training and pedestrian target in grader for use.

Gradient calculation is specific as follows：

Gradient in gray level image each pixel two-dimensional coordinate plane is calculated, formula is as follows：

Grad_x(x, y)=IM (x+1, y)-IM (x-1, y)

Grad_y(x, y)=IM (x, y+1)-IM (x, y-1),

Wherein, Grad_x、Grad_yWith the gradient that IM (x, y) is respectively abscissa, the gradient on ordinate, image intensity value Size, therefore, the gradient calculation of each pixel is as follows in image：

Wherein, Grad (x, y) is the gradient magnitude size of IM (x, y), and θ (x, y) is that IM (x, y) gradient direction angle is big It is small.

Adjacent scale feature approximate calculation is as follows：

1) for source images Im (x, y) for giving, the factor of influence τ of picture up-sampling, by the image of up-sampling by Im' (x, y) is represented, the formula of conversion is as follows：

Im'(x, y)=Im (x/ τ, y/ τ)；

2) mode calculated according to image gradient can be seen that the gradient of image gradient information and source images after up-sampling There is following relation in information：

3) relation therefore between the total gradient magnitude of up-sampling image and the gradient magnitude of source images is as follows：

Under general environment and complex environment, time consuming comparing is calculated using HOG, HogLbp and the inventive method, Result is as shown in Table 1 and Table 2.

Pedestrian detection method time contrast table under the general environment of table 1

Pedestrian detection method time contrast table under the complex environment of table 2

As can be seen from the above table in the case where swift nature approximate calculation is not used, the time of first two method consumption is very Many, the detection time that the method for the present invention spends is minimum.As can be seen here, using the method for swift nature approximate calculation, in not shadow The speed of detection can be improved in the case of ringing Detection results, contributes to real-time to detect pedestrian target.

Claims (4)

1. a kind of pedestrian detection method, it is characterised in that comprise the following steps：

S1：Source images are pre-processed by collection pedestrian's video image；

S2：The primary detection of pedestrian：

S2-1：Extract great amount of images and be used as training sample, positive sample is had to comprising pedestrian, and negative sample is various not comprising pedestrian The combination of background picture；

S2-2：Feature extraction is carried out with improved HOG feature extracting methods；

S2-3：A two-value grader is built according to characteristic is extracted, sample training is carried out to grader；

S2-4：Extract great amount of images and be used as detection sample, feature extraction is carried out with improved HOG feature extracting methods；

S2-5：It is identified by grader, obtains primary pedestrian detection result；

S3：Pedestrian's accurate detection：

S3-1：Extract great amount of images and be used as training sample, positive sample is had to comprising pedestrian, and negative sample is various not comprising pedestrian The combination of background picture；

S3-2：The extraction of color, gradient, histogram feature is carried out using Integration tunnel feature extracting method, using adjacent yardstick Feature approximate calculation method calculates adjacent scale feature；

S3-3：It is trained using Adaboost obtains final strong classifier for target classification；

S3-4：The primary pedestrian detection result images obtained to step S2, carry out Integration tunnel feature extracting method and extract feature, Grader using step S3-3 is classified, and obtains accurate pedestrian detection result.

2. pedestrian detection method according to claim 1, it is characterised in that the improved HOG feature extracting methods tool Body is as follows：

1) source images are carried out into greyscale transformation and obtains corresponding gray level image；

2) Gamma corrections i.e. color space normalized is carried out to the gray level image for obtaining；

3) gradient of each pixel is calculated, the profile information of target image is obtained；

4) N number of Cell (cell compartment) is divided an image into, then several adjacent Cs ell is combined into a Block (square area Domain), the histogram of gradients in each Cell is calculated, the Feature Descriptor connection that all of Cell in Block is obtained is arrived every The feature of individual Block；

5) it is as follows histogram of gradients as the probability density function in comentropy definition：

$E=-\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\θ}}_{i}log\left({\mathrm{\θ}}_i\right),$

The θ of formula_iThe ratio of the gradient number and total gradient number in a gradient direction interval is represented, gradient direction in Block is calculated The interval manhole ladder number of degrees and the ratio of total gradient number, obtain the information entropy of each block, are used for comentropy as sample characteristics The training of grader and the detection of pedestrian target.

3. pedestrian detection method according to claim 1, it is characterised in that the gradient calculation is specific as follows：

Gradient in gray level image each pixel two-dimensional coordinate plane is calculated, formula is as follows：

$\begin{array}{c}{\mathrm{Grad}}_x(x,y)=IM(x+1,y)-IM(x-1,y)\\ {\mathrm{Grad}}_y(x,y)=IM(x,y+1)-IM(x,y-1)\end{array},$

Wherein, Grad_x、Grad_yWith the gradient that IM (x, y) is respectively abscissa, the gradient on ordinate, image intensity value size, Therefore, the gradient calculation of each pixel is as follows in image：

$Grad(x,y)=\sqrt{{\mathrm{Gard}}_x{(x,y)}^2+{\mathrm{Gard}}_y{(x,y)}^2},$

$\mathrm{\θ}(x,y)={\tan }^{-1}\left(\frac{{\mathrm{Gard}}_y(x,y)}{{\mathrm{Gard}}_x(x,y)}\right)$

Wherein, Grad (x, y) is the gradient magnitude size of IM (x, y), and θ (x, y) is IM (x, y) gradient direction angle size.

4. pedestrian detection method according to claim 1, it is characterised in that the adjacent scale feature approximate calculation is such as Under：

1) for source images Im (x, y) for giving, the factor of influence τ of picture up-sampling, by the image of up-sampling by Im'(x, Y) represent, the formula of conversion is as follows：

Im'(x, y)=Im (x/ τ, y/ τ)；

2) mode calculated according to image gradient can be seen that the gradient information of image gradient information and source images after up-sampling In the presence of following relation：

$Grad{\mathrm{Im}}^{\′}(x,y)\≈\frac{1}{\mathrm{\τ}}Grad\mathrm{Im}(x/\mathrm{\τ},y/\mathrm{\τ});$

3) relation therefore between the total gradient magnitude of up-sampling image and the gradient magnitude of source images is as follows：

$\begin{array}{c}\underset{i=1}{\overset{\mathrm{\τ}l}{\mathrm{\Σ}}}\underset{j=1}{\overset{\mathrm{\τ}k}{\mathrm{\Σ}}}Grad{\mathrm{Im}}^{\′}\left(x,y\right)\≈\underset{i=1}{\overset{\mathrm{\τ}l}{\mathrm{\Σ}}}\underset{j=1}{\overset{\mathrm{\τ}k}{\mathrm{\Σ}}}1/\mathrm{\τ}Grad\mathrm{Im}\left(x,y\right)\\ \≈{\mathrm{\τ}}^2\underset{i=1}{\overset{l}{\mathrm{\Σ}}}\underset{j=1}{\overset{k}{\mathrm{\Σ}}}1/\mathrm{\τ}Grad\mathrm{Im}\left(x,y\right)\\ \≈\mathrm{\τ}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}\underset{j=1}{\overset{k}{\mathrm{\Σ}}}Grad\mathrm{Im}\left(x,y\right)\end{array}.$