Method and apparatus for detecting persons, and non-transitory computer-readable recording medium

1. A method for detecting persons, the method comprising:

initially detecting the persons in a height-top-view including the persons to be detected;

dividing the height-top-view into one or more regions, and estimating crowd density in each of the regions;

determining, based on the crowd density, visible regions of the initially detected persons in each of the regions;

for each of the initially detected persons, extracting a first gradient feature of the person from the height-top-view, and a second gradient feature of the person from a grayscale image or a color image corresponding to the height-top-view, respectively;

for each of the initially detected persons, determining, based on the extracted first gradient feature and second gradient feature, using a previously constructed classifier corresponding to the determined visible region of the person, a confidence level of the initially detected person; and

correcting, based on the confidence level, a detection result of the initially detected persons.

2. The method for detecting persons according to claim 1 ,

wherein dividing the height-top-view into one or more regions and estimating the crowd density in each of the regions further includes

performing clustering for the persons initially detected in the height-top-view to form one or more clustering regions, and

for each of the clustering regions, calculating, based on a number of the persons initially detected in the clustering region and area of the clustering region, the crowd density.

3. The method for detecting persons according to claim 1 ,

wherein dividing the height-top-view into one or more regions and estimating the crowd density in each of the regions further includes

dividing the height-top-view into regions with a predetermined size,

for each of the regions, extracting a density detection feature, the density detection feature indicating distribution of differences in height between pixels in the region,

detecting, based on the density detection feature, using a density model previously constructed by a statistical learning method, a number of persons in each of the regions, and

for each of the regions, calculating, based on the detected number of the persons in the region and area of the region, the crowd density.

4. The method for detecting persons according to claim 1 ,

wherein determining the visible regions of the initially detected persons in each of the regions based on the crowd density further includes, for each of the regions

comparing the crowd density of the region with a first threshold and a second threshold, the first threshold being greater than the second threshold,

determining that the visible regions of the persons initially detected in the region are head-regions, when the crowd density is greater than or equal to the first threshold,

determining that the visible regions of the persons initially detected in the region are head-shoulder-regions, when the crowd density is less than the first threshold and is greater than or equal to the second threshold, and

determining that the visible regions of the persons initially detected in the region are complete regions, when the crowd density is less than the second threshold.

5. The method for detecting persons according to claim 1 ,

wherein for each of the initially detected persons, extracting the first gradient feature of the person from the height-top-view, and the second gradient feature of the person from the grayscale image or the color image corresponding to the height-top-view, respectively includes

extracting the first gradient feature from a region, which includes the initially detected person, of the height-top-view,

determining, based on a position and a pixel value of a highest point of the initially detected person in the height-top-view, by coordinate transformation, a position of the initially detected person in the corresponding grayscale image or color image, and a circumscribed rectangle of the complete and initially detected person,

determining a portion, which corresponds to the visible region of the initially detected person, of the circumscribed rectangle, and

extracting, from the corresponding portion of the circumscribed rectangle in the corresponding grayscale image or color image, the second gradient feature.

6. The method for detecting persons according to claim 5 ,

wherein the first gradient feature is a LBP (Local Binary Pattern) histogram, and the second gradient feature is a HOG (Histogram of Oriented Gradients) feature.

7. The method for detecting persons according to claim 1 ,

wherein the classifier is obtained by training using the first gradient feature and the second gradient feature, and the classifier includes a classifier corresponding to a head-region of a person, a classifier corresponding to a head-shoulder-region of a person, and a classifier corresponding to a complete region of a person.

8. The method for detecting persons according to claim 1 ,

wherein correcting the detection result of each of the initially detected persons based on the confidence level includes

comparing the confidence level of the initially detected person with a confidence level threshold,

determining that the initially detected person is the person to be detected, when the confidence level is greater than or equal to the confidence level threshold, and

determining that the initially detected person is a false detection, when the confidence level is less than the confidence level threshold.

9. The method for detecting persons according to claim 3 ,

wherein correcting the detection result of the initially detected persons based on the confidence level includes

for each of the regions, obtaining a solution of an optimization equation

min

x

i

∈

{

0

,

1

}

⁢

(

D

-



X



-

X

′

×

p

⁡

(

X

)

)

to obtain the corrected detection result of the persons,

where D represents the crowd density in the region, X is vector representation of the person initially detected in the region, X i is an i-th element in the vector X, and p(X) is a confidence level vector of the person initially detected in the region.

10. An apparatus for detecting persons, the apparatus comprising:

an initial-detection unit configured to initially detect the persons in a height-top-view including the persons to be detected;

a density estimation unit configured to divide the height-top-view into one or more regions, and estimate crowd density in each of the regions;

a visible region determination unit configured to determine, based on the crowd density, visible regions of the initially detected persons in each of the regions;

a feature extraction unit configured to extract, for each of the initially detected persons, a first gradient feature of the person from the height-top-view and a second gradient feature of the person from a grayscale image or a color image corresponding to the height-top-view, respectively;

a confidence level determination unit configured to determine, for each of the initially detected persons, based on the extracted first gradient feature and second gradient feature, using a previously constructed classifier corresponding to the determined visible region of the person, a confidence level of the initially detected person; and

a correction unit configured to correct, based on the confidence level, a detection result of the initially detected persons.

11. A non-transitory computer-readable recording medium having stored therein a program for causing a computer to execute a process for detecting persons, the process comprising:

initially detecting the persons in a height-top-view including the persons to be detected;

dividing the height-top-view into one or more regions, and estimating crowd density in each of the regions;

determining, based on the crowd density, visible regions of the initially detected persons in each of the regions;

for each of the initially detected persons, extracting a first gradient feature of the person from the height-top-view, and a second gradient feature of the person from a grayscale image or a color image corresponding to the height-top-view, respectively;

for each of the initially detected persons, determining, based on the extracted first gradient feature and second gradient feature, using a previously constructed classifier corresponding to the determined visible region of the person, a confidence level of the initially detected person; and

correcting, based on the confidence level, a detection result of the initially detected persons.