CN102799894B - Detection method for rapidly overlooking pedestrians - Google Patents

Abstract

The invention discloses a detection method for rapidly overlooking pedestrians. The method comprises the following steps of: overlooking the pedestrians by using an overhead monocular camera, setting multiple stages of classifiers by utilizing the geometric characteristics of heads or heads and shoulders of the pedestrians and combining the characteristic that the colors of the heads are uniform, and setting rectangular frame areas which frame the heads in images as output targets of the classifiers; calculating accumulated values of gray levels of all pixels in rectangular subblocks of the classifiers by using an integrogram, calculating the deviations of the accumulated values in rectangular frames of the classifiers, and judging and screening; and traversing the whole images by using rectangular windows with set sizes, circularly calculating the accumulated values of the gray levels of the rectangular subblocks of the classifiers and the deviations of the accumulated values between the corresponding rectangular subblocks for all candidate areas, and rapidly capturing head portrait information of the overlooked pedestrians. The characteristic operation which is involved in the method is mainly an additive operation, and the cascade design of the classifiers can ensure that most of non-target areas are filtered when passing through the previous classifier, so that the calculated amount is greatly reduced, and the speed of capturing the head portrait information of the pedestrians is improved.

Description

A kind of detection method of overlooking pedestrian fast

Technical field

The present invention relates to a kind of detection method of overlooking pedestrian fast, for stream of people's highly dense places such as large-scale exhibition center, stadium, library, airport, subway station, bus, markets, to the statistics of passenger flow.

Background technology

China is a populous country, and especially in city, the rational allocation of public transport and public resource is the problem that relevant functional department shows great attention to always.In stream of people's highly dense places such as large-scale exhibition center, stadium, library, airport, subway station, bus, markets, the statistics of passenger flow is seemed particularly important, only reliable passenger flow data accurately, the effect of competence exertion intelligent dispatching system, allow public resource be utilized more fully, allow the more convenient and safety of popular trip.

At present according to data acquisition principle, passenger flow counting system can be divided into infrared photoelectric sensor formula, multi-cam stereo vision formula and single camera formula three major types system.

1, infrared photoelectric sensor formula passenger flow counting system adopts infrared emission to block and pressure sensor technique, this system effectively can distinguish the interference of adult in passenger flow and child and luggage and articles, but crowded crowd cannot be processed, inapplicable common bus passenger flow detects, and can only be used in the single detection and stenostomatous coach bus work of coming in and going out overloads.

2, the passenger flow counting system of the stereoscopic vision of multi-cam.Use at least two cameras demarcated in advance, by three-dimensional depth algorithm, the depth map of scene is calculated.Each pedestrian before camera can be partitioned in depth map.The subject matter of multi-cam stereo vision formula system is owing to using multiple camera, and needs accurate calibration, and volume is comparatively large, and system cost is also higher.

The people such as Terada propose to utilize binocular stereo vision to obtain the depth information of pedestrian, and detect target according to depth information; (" A Method of Counting the Passing People by Using the Stereo Images; " In Proc.Intl.Conf.of Image Processing, Vol.2, pp.338-342, Oct, 1999) people proposes based on pedestrian head feature and utilizes the parallax of head zone to realize motion people counting in beach etc.; The people such as (journal of Zhejiang university (engineering version) 2009.14 (3) .) Zhu Qiuyu propose a kind of bus passenger flow method of counting based on stereoscopic vision (Journal of Image and Graphics, 2009.14 (11): the 2391-2395 pages); But the problems such as calculated amount is excessive, the reliability of motion tracking is not high that the above-mentioned method based on stereoscopic vision exists.

3, the passenger flow counting system of monocular cam.The method of counting of monocular cam before this, mainly detects the number of people or head shoulder, counts according to shape, area or colour recognition passenger.Such as, in beach, Liu Jilin is at paper: the machine vision method being applied to bus passenger flow statistics, the hough change detection proposed in (Journal of Image and Graphics, 2008.13 (4): the 716-722 pages) circle or class circle; Paper: K-means Based Segmentation for Real-Time Zenithal People Counting, the method of (in IEEE International Conference on Image Processing.2008.) first carries out background removal, the object block of separation prospect, by experiment object block is divided into the 8x8 of standard, 12x12,16x16 tri-kinds of specifications, and set the minor increment of block and block, then use k-means to obtain the quantity of pedestrian to foreground target block comminute; Paper [6] is the strategy based on region, namely by the magnitude estimation pedestrian quantity of zoning pixel.These methods have plenty of the specific features not detecting target, and have plenty of the minutia excessively relying on target, as rim detection, colouring information etc., these are placed on the operational effect under the actual operating conditions of complexity, need inspection.

Summary of the invention

The invention discloses a kind of detection method of overlooking pedestrian fast, utilize the geometry symmetry of head or the head shoulder overlooking pedestrian or asymmetric feature, in conjunction with the uniform feature of head colour stable, the gray scale accumulated value of image rectangular area is compared, to reach the object detecting pedestrian fast.The inventive method can effectively overcome in prior art, and infrared photoelectric sensor cannot detect the crowded stream of people; Multi-cam stereo visual method volume is comparatively large, and system cost is high, and calculated amount is large, and the reliability of motion tracking is not high; Monocular cam method is due to operating condition complexity, and effect there is no the drawbacks such as method inspection.

The detection method that the present invention overlooks pedestrian is fast achieved in that

Overlook a pedestrian's detection method fast, overlooked image or the video of pedestrian by the monocular cam shooting of overhead, utilize the geometrical feature of pedestrian head, in conjunction with the uniform feature of head color, the sorter structure of setting multi-stage cascade; The rectangle frame region confining head is set to the output target of sorter; Utilize " integrogram " to calculate the gray scale accumulated value of each sorter rectangular sub blocks, according to characteristic of correspondence, calculate the deviation of each sorter rectangle frame inner rectangular sub-block accumulated value, carry out judging, screening; Whole image is traveled through with the rectangular window setting size, to whole image each confine region successively, the rectangular sub blocks gray scale accumulated value of each sorter requirement of cycle calculations and deviation, require just to be abandoned at once once not meet sorter grade in inspection region, continue again to process follow-up region to be checked, thus Quick Catch overlooks the head image information of pedestrian.

Setting and the concrete implementation step of the sorter structure of described multi-stage cascade are as follows:

A) sorter I: head rectangle frame is considered as bilateral symmetry, the region confined by rectangular window is divided into two by left and right, become two rectangular sub blocks, " integrogram " is utilized to calculate the gray scale accumulated value of left and right two rectangular sub blocks respectively, recycling formula (3) calculates the gray-scale value deviation in left and right two rectangular areas, and the calculating formula of deviation e is as follows:

$e=\left|\frac{\mathrm{RecSum}\left(l\right)-\mathrm{RecSum}\left(r\right)}{\max (\mathrm{RecSum}\left(l\right),\mathrm{RecSum}\left(r\right))}\right|---\left(3\right)$

In formula, RecSum (l), RecSum (r) represent the gray-scale value sum of all pixels in left and right two rectangular blocks respectively; E deviate allowed band, 0.005 ~ 0.2, requires just to enter step B if calculating gained deviation meets setting value), otherwise mobile rectangular window enters steps A);

B) sorter II: head rectangle frame is divided into left, center, right structure, the region confined by rectangular window is divided into impartial three pieces from left to right, and wherein the pixel gray scale accumulated value of three pieces is respectively from left to right: RecSum (2), RecSum (l), RecSum (3); According to the ellipse of the number of people or the feature of sub-circular, if the gray scale accumulated value Recsum (1) of pixel is greater than the RecSum (2) on the left side and the gray scale accumulated value RecSum (3) of right pixels point respectively in intermediate rectangular region, then enter step C), otherwise mobile rectangular window enters steps A);

C) sorter III: head rectangle frame is considered as upper and lower equal, the region confined by rectangular window is divided into two by upper and lower, become two rectangular sub blocks, " integrogram " is utilized to calculate the gray scale accumulated value of upper and lower two rectangular sub blocks respectively, formula (3) is utilized to calculate gray-scale value deviation in upper and lower two rectangular areas, wherein: RecSum (t), RecSum (b) represent the gray-scale value sum of all pixels in upper and lower two rectangular sub blocks respectively;

E deviate allowed band, 0.005 ~ 0.2, requires just to enter step D if calculating gained deviation meets setting value), otherwise mobile rectangular window enters steps A);

D) sorter IV: utilize " integrogram " to calculate the core region 30x30 pixel of head original image rectangle region, and confine rectangle frame;

E) mobile rectangle frame circulation step A) ~ D), travel through whole image with the rectangular window of 50x50 pixel size, obtain target complete rectangle frame;

F) size detecting rectangular window is reset by different scale, on the basis of initial 50x50 window size, length and width respectively expand 10 pixels, namely the window of 60x60,70x70,80x80 is used again according to steps A)-F) calculate, until finally confirm and confine head target, the setting of concrete yardstick and window size can adjust according to actual image information;

G) utilize non-maximum suppression principle, all rectangle frames confining same target are screened, curb the rectangle frame of those overlapping area more than 10%, retain the rectangle frame that the variance of central area gray-scale value is minimum.

Described monocular cam is arranged on upper end, top, public passage gateway, takes pedestrian downwards, and camera is apart from 3 ~ 10 meters, ground.

The present invention utilizes the geometrical feature of pedestrian head or head shoulder, in conjunction with the uniform feature of head color, the head zone of overlooking pedestrian is represented with rectangle frame, is become by the character representation of head: rectangle inside is left and right, and to divide equally the gray scale accumulated value of two halves equal; Rectangle inside is upper and lower, and to divide equally the gray scale accumulated value of two halves equal; Rectangle inside left, center, right three part, the gray scale accumulated value of center section is greater than the value of both sides part; Rectangular centre district gray-scale value is even.Utilize " integrogram ", the gray scale accumulated value of any rectangle frame area pixel in computed image, the uniformity coefficient of gray-scale value represents by the variance of gray-scale value in region.

Arrange the sorter structure of plural serial stage, four characteristic Design of the above-mentioned number of people pressed by sorter, the corresponding Weak Classifier of each feature, therefore the sorter of system is in series by 4 Weak Classifiers.This makes most of nontarget area through what sorter being above just filtered out, and greatly reduces calculated amount, improves the speed catching and overlook pedestrian's head image information.According to different features, the gray scale accumulated value of inner for rectangle frame formed objects, zones of different is subtracted each other gained difference, as the eigenwert that sorter exports, can be used for learning, or carry out classification judgement.

Adopt the rectangle frame of different scale size, travel through whole image as detection window, catch because height differences causes the head image information of different size.

The quantity of sorter can also suitably increase according to actual conditions or reduce.

Advantage of the present invention and good effect are:

The characteristic operation related to inside the present invention is all signed magnitude arithmetic(al) substantially, and sorter is cascade structure, the target only having only a few real is made just to need through all detection of classifiers, major part nontarget area is just being filtered out through previous classifiers, greatly reduce calculated amount, improve the speed catching pedestrian's head image information.

Accompanying drawing explanation

Fig. 1 is the structural representation of the present invention's four kinds of sorters;

Fig. 2 is the pixel in integrogram;

Fig. 3 is the rectangular area in integrogram.

A): sorter I, b): sorter II, c): sorter III, d): sorter IV.

Embodiment

Below in conjunction with drawings and Examples, the present invention is described in detail.

A kind of detection method of overlooking pedestrian fast of the present invention, different from traditional dependence target detail feature, the present invention does not obtain the information such as the edge of target, shape and color, and the space geometry characteristic of just captured target shape; Overlooked image or the video of pedestrian by the monocular cam of overhead, utilize the geometrical feature of pedestrian head, in conjunction with the uniform feature of head color, the sorter structure of setting multi-stage cascade; The rectangle frame region confining head is set to the output target of sorter; Utilize " integrogram " to calculate the gray scale accumulated value of each sorter rectangular sub blocks, according to characteristic of correspondence, calculate the deviation of each sorter rectangle frame inner rectangular sub-block accumulated value, carry out judging, screening; Whole image is traveled through with the rectangular window setting size, to whole image each confine region successively, the rectangular sub blocks gray scale accumulated value of each sorter requirement of cycle calculations and deviation, require just to be abandoned at once once not meet sorter grade in inspection region, continue again to process follow-up region to be checked, thus Quick Catch overlooks the head image information of pedestrian.

Setting and the concrete implementation step of the sorter structure of described multi-stage cascade are as follows:

A) sorter I: head rectangle frame is considered as bilateral symmetry, the region confined by rectangular window is divided into two by left and right, become two rectangular sub blocks, " integrogram " is utilized to calculate the gray scale accumulated value of left and right two rectangular sub blocks respectively, recycling formula (3) calculates the gray-scale value deviation in left and right two rectangular areas, and the calculating formula of deviation e is as follows:

$e=\left|\frac{\mathrm{RecSum}\left(l\right)-\mathrm{RecSum}\left(r\right)}{\max (\mathrm{RecSum}\left(l\right),\mathrm{RecSum}\left(r\right))}\right|---\left(3\right)$

In formula, RecSum (l), RecSum (r) represent the gray-scale value sum of all pixels in left and right two rectangular blocks respectively;

E deviate allowed band, 0.005 ~ 0.2, requires just to enter step B if calculating gained deviation meets setting value), otherwise mobile rectangular window enters steps A);

B) sorter II: head rectangle frame is divided into left, center, right structure, the region confined by rectangular window is divided into impartial three pieces from left to right, and wherein the pixel gray scale accumulated value of three pieces is respectively from left to right: RecSum (2), RecSum (l), RecSum (3); According to the ellipse of the number of people or the feature of sub-circular, if the gray scale accumulated value Recsum (1) of pixel is greater than the RecSum (2) on the left side and the gray scale accumulated value RecSum (3) of right pixels point respectively in intermediate rectangular region, then enter step C), otherwise mobile rectangular window enters steps A);

C) sorter III: head rectangle frame is considered as upper and lower equal, the region confined by rectangular window is divided into two by upper and lower, become two rectangular sub blocks, " integrogram " is utilized to calculate the gray scale accumulated value of upper and lower two rectangular sub blocks respectively, formula (3) is utilized to calculate gray-scale value deviation in upper and lower two rectangular areas, wherein: RecSum (t), RecSum (b) represent the gray-scale value sum of all pixels in upper and lower two rectangular sub blocks respectively;

E deviate allowed band, 0.005 ~ 0.2, requires just to enter step D if calculating gained deviation meets setting value), otherwise mobile rectangular window enters steps A);

D) sorter IV: utilize " integrogram " to calculate the core region 30x30 pixel of head original image rectangle region, and confine rectangle frame;

E) mobile rectangle frame circulation step A) ~ D), travel through whole image with the rectangular window of 50x50 pixel size, obtain target complete rectangle frame;

F) size detecting rectangular window is reset by different scale, on the basis of initial 50x50 window size, length and width respectively expand 10 pixels, namely the window of 60x60,70x70,80x80 is used again according to steps A) F) calculate, until finally confirm and confine head target, the setting of concrete yardstick and window size can adjust according to actual image information;

G) utilize non-maximum suppression principle, all rectangle frames confining same target are screened, curb the rectangle frame of those overlapping area more than 10%, retain the rectangle frame that the variance of central area gray-scale value is minimum.

Described monocular cam is arranged on upper end, top, public passage gateway, takes pedestrian downwards, and camera is apart from 3 ~ 10 meters, ground.

The present invention does not directly use black as number of people clarification of objective, but is determined by the variance calculating all pixel values in rectangle frame.That is, the metastable region of gray-scale value is only had may to be just head.Like this, to those hair-dyeings or the pedestrian of bald head also can effectively identification.

These features are designed to simple sorter by respectively, it is the relation of cascade between sorter, that is, all target level by level to be detected enter detection of classifier, the sorter not meeting any one-level all can be abandoned halfway, only have meet characteristic target be only our real target.

Described " integrogram " is exactly a kind of " cumulative sum table ", and namely " integrogram " add up line by line on original image gray-scale value of each point pixel obtains, as shown in Figure 2.In integrogram, the expression formula of any pixel is:

In formula, i represents the value (conventional is gray-scale value) of original image pixel, and ii represents the value of this pixel in integral image.Visible, the value of the arbitrary pixel in integral image, equals the cumulative of the value of all pixels in rectangular area, upper left of this position of original image, i.e. integrated value.

If solve the integrated value of all pixels in any rectangle in integrogram, as the integrated value of Tu3Zhong D district pixel, then have

ii(D)=ii(4)+ii(1)–ii(2)–ii(3)（2）

Ii (D) represents the integrated value of rectangle region D in integrogram.Above formula shows, the integrated value of any rectangle region in integrogram, is determined by the integrated value plus and minus calculation of this rectangle four points.Thus, can the integrated value of any rectangular area of rapid solving one sub-picture.Overlook the pedestrian detection technology of passenger flow herein, the characteristic operation that the inside relates to is mainly calculated on the basis of integrogram.

Claims (2)

1. overlook a pedestrian's detection method fast, overlooked image or the video of pedestrian by the monocular cam shooting of overhead, utilize the geometrical feature of pedestrian head, in conjunction with the uniform feature of head color, the sorter structure of setting multi-stage cascade; It is characterized in that: setting and the concrete implementation step of the sorter structure of described multi-stage cascade are as follows:

A) sorter I: head rectangle frame is considered as bilateral symmetry, the region confined by rectangular window is divided into two by left and right, become two rectangular sub blocks, " integrogram " is utilized to calculate the gray scale accumulated value of left and right two rectangular sub blocks respectively, recycling formula (3) calculates the gray-scale value deviation in left and right two rectangular areas, and the calculating formula of deviation e is as follows:

$e=\left|\frac{\mathrm{RecSum}\left(1\right)-\mathrm{RecSum}\left(r\right)}{\max (\mathrm{RecSum}\left(1\right),\mathrm{RecSum}\left(r\right))}\right|---\left(3\right)$

In formula, RecSum (l), RecSum (r) represent the gray-scale value sum of all pixels in left and right two rectangular blocks respectively;

E deviate allowed band, 0.005 ~ 0.2, requires just to enter step B if calculating gained deviation meets setting value), otherwise mobile rectangular window enters steps A);

B) sorter II: head rectangle frame is divided into left, center, right structure, the region confined by rectangular window is divided into impartial three pieces from left to right, and wherein the pixel gray scale accumulated value of three pieces is respectively from left to right: RecSum (2), RecSum (l), RecSum (3); According to the ellipse of the number of people or the feature of sub-circular, if the gray scale accumulated value Recsum (1) of pixel is greater than the RecSum (2) on the left side and the gray scale accumulated value RecSum (3) of right pixels point respectively in intermediate rectangular region, then enter step C), otherwise mobile rectangular window enters steps A);

C) sorter III: head rectangle frame is considered as upper and lower equal, the region confined by rectangular window is divided into two by upper and lower, become two rectangular sub blocks, " integrogram " is utilized to calculate the gray scale accumulated value of upper and lower two rectangular sub blocks respectively, formula (3) is utilized to calculate gray-scale value deviation in upper and lower two rectangular areas, wherein: RecSum (t), RecSum (b) represent the gray-scale value sum of all pixels in upper and lower two rectangular sub blocks respectively;

E deviate allowed band, 0.005 ~ 0.2, requires just to enter step D if calculating gained deviation meets setting value), otherwise mobile rectangular window enters steps A);

D) sorter IV: utilize " integrogram " to calculate the core region 30x30 pixel of head original image rectangle region, and confine rectangle frame;

E) mobile rectangle frame circulation step A) ~ D), travel through whole image with the rectangular window of 50x50 pixel size, obtain target complete rectangle frame;

F) size detecting rectangular window is reset by different scale, on the basis of initial 50x50 window size, length and width respectively expand 10 pixels, namely the window of 60x60,70x70,80x80 is used again according to steps A)-F) calculate, until finally confirm and confine head target, the setting of concrete yardstick and window size can adjust according to actual image information;

G) utilize non-maximum suppression principle, all rectangle frames confining same target are screened, curb the rectangle frame of those overlapping area more than 10%, retain the rectangle frame that the variance of central area gray-scale value is minimum.

2. a kind of detection method of overlooking pedestrian fast according to claim 1, is characterized in that: described monocular cam is arranged on upper end, top, public passage gateway, takes pedestrian downwards, and camera is apart from 3 ~ 10 meters, ground.