CN103150550B - A kind of road pedestrian event detection method based on gripper path analysis - Google Patents

Abstract

The invention provides a kind of road pedestrian event detection method based on gripper path analysis, by being divided into each piece in multiple pieces of area images, the background block identical with this block position is found in background image, calculate the absolute value sum of gray scale difference value and assignment, determine object block, find best angle point, obtain unique point, create an object construction body simultaneously, record these clarification of objective point position and matched jamming counter informations, search in current frame image by template, repeat said process, obtain the pursuit path of target, by searching mapping table, obtain the actual range that pursuit path is corresponding, try to achieve the speed of pursuit path, judge whether this target is pedestrian.Detection method of the present invention, can detect pedestrian targets all in range of video, not by environmental restraint, can detect real-time video, and detection time short, be easy to realize, accuracy is higher, be well suited for real-time detect lines people event, have broad application prospects.

Description

A kind of road pedestrian event detection method based on gripper path analysis

Technical field

The invention belongs to field of video detection, be specifically related to a kind of road pedestrian event detection method based on gripper path analysis.

Background technology

Road pedestrian event refers to that pedestrian enters without any when safeguard measure on car lane, the behavior that jammer motor-car normally travels.Although traffic control department takes measures, the situation that pedestrian swarms into car lane happens occasionally, and its danger is very large, easily causes traffic congestion, even leads to traffic hazard, cause serious impact to people's normal life.Traditional pedestrian event detection method mainly contains temperature checking method, electronic coil detection method, digital video detection method, and wherein temperature checking method is easily subject to vehicle interference; Electronic coil poor expandability, must suspend traffic during installation and maintenance, destroy road surface, and these methods can not be used widely in real life.

Current new project adopts installation more and more, safeguard do not need to destroy roadbed, surveyed area large, implement the convenient, flexible transport information detection technique based on video.Pedestrian detection method based on video becomes the focus of research, and existing method mainly contains based on neural network pedestrian detection, based on the template matches detection method etc. of wavelet transformation.Although these methods can realize pedestrian's affair alarm, the complex disposal process of video data, poor reliability, can not meet the requirement of real-time of detection, cannot meet the requirement of practical application.

Summary of the invention

For shortcomings and deficiencies of the prior art, the object of the invention is to, provide a kind of road pedestrian event detection method based on gripper path analysis, the method can realize detecting in real time, reliably to pedestrian's events all in range of video.

In order to realize above-mentioned task, the present invention adopts following technical scheme to be achieved:

Based on a road pedestrian event detection method for gripper path analysis, the method is carried out according to following steps:

Step one, sets up the mapping relations of image pixel to road surface actual range, i.e. mapping table;

Step 2, first two field picture is all divided into multiple pieces of regions with background image under identical block coordinate system, and the size of background is W*H, and the block size of division is w*h, the block areal T divided is T=(W/w) * (H/h), all pixel value B in i-th block _irepresent, total number N=w*h, the B of pixel in i-th block _imiddle preservation w*h pixel value, with the lower left corner of i-th block for initial point sets up right angle two-dimensional coordinate system Y, the pixel value B that in i-th block, (m, n) puts _i(m, n) represents, wherein:

W is the pixel in background level direction;

H is the pixel of background vertical direction;

W is the pixel of the Width of i-th block;

H is the pixel of the short transverse of i-th block;

i＝1,2,3...T；

M represents the horizontal ordinate of arbitrary pixel under coordinate system Y in i-th block, m=0,1,2...w-1;

N represents the ordinate of arbitrary pixel under coordinate system Y in i-th block, n=0,1,2...h-1;

Step 3, to each piece in the first two field picture, finds the background block identical with this block position in background image, and calculates the absolute value sum of the gray scale difference value of each same pixel position between this block its corresponding background block;

When the absolute value sum of gained is greater than the threshold value A of setting, then this block is object block, is 255 by the gray-scale value assignment of pixels all in this object block,

When the absolute value sum of gained is less than or equal to the threshold value A of setting, then this block is background block, is 0 by the gray-scale value assignment of pixels all in background block, wherein:

The span of described threshold value A is the area of (10 ~ 20) × block;

Finally the background in the first two field picture and target are separated, obtain the binary image of the first two field picture;

Step 4, to the binary image obtained according to from left to right, order from top to bottom scans in units of block, and adjacent object block is labeled as same target, calculates height and the width of each target-marking simultaneously, when the value of height/width is within the scope of threshold value B, rim detection is carried out to this binary image, finds best angle point, namely when laterally detect data with longitudinally detect data be greater than threshold value T simultaneously time, judge that this point is as best angle point, wherein:

The scope of described threshold value B is 2 ~ 10;

The value of described threshold value T is 180;

Step 5, to select in these angle points laterally to detect data and longitudinally detect data and minimum point as clarification of objective point, create an object construction body simultaneously, record these clarification of objective point position and matched jamming counter informations, matched jamming counter R first time is initialized as zero, and unique point coordinate is (x ₁, y ₁), record the image of a block size centered by this unique point as template, the pixel value two-dimensional array B of all pixels in this template simultaneously _t[N] represents, in this block, the pixel value of pixel (m, n) is B _t(m, n);

Step 6, in the second two field picture, with the image block B of the first frame recording _t[N] is template, with the characteristic point position (x of this template ₁, y ₁) centered by, the square area choosing 4 block sizes in current frame image is as region of search, and according to from left to right, method from top to bottom, searches for one by one, and the number of to be searched piece is designated as N, uses B _j[N] represents any one the search block in current search region, obtains both definitely value difference and SAD,

Wherein:

$\mathrm{SAD}=\underset{m=0}{\overset{w-1}{\mathrm{\Σ}}}\underset{n=0}{\overset{h-1}{\mathrm{\Σ}}}|B_j(m,n)-B_t(m,n)|$

j＝1,2,3....N

Minimum as matching criterior using absolute length chang, choose N number of absolute value difference and in minimum block be match block, be designated as B _s[N], namely have found matching characteristic point in the current frame, records the position (x of new matching characteristic point simultaneously ₂, y ₂), and in order to the block image B centered by new unique point _s[N] is as new template B _t[N], matched jamming counter R adds 1 simultaneously;

Step 7, from the 3rd two field picture to M two field picture, M be greater than 60 positive integer, repeat step 6, when R equals threshold value C, perform step 8;

Step 8, when matched jamming counter R equals threshold value C, all unique point (x ₁, y ₁) ... (x ₆₀, y ₆₀) form the pursuit path of target, by the mapping table set up in finding step one, obtain the actual range that pursuit path is corresponding, consecutive point be spaced apart every frame time, so can obtain the array of one group of time and actual range, utilize least square fitting can try to achieve the speed of pursuit path, when this speed meets the scope of threshold value D, namely determine that this target is for pedestrian, wherein:

The value of described threshold value C is 60;

The scope of described threshold value D is (0.3 ~ 2.0) m/s.

Road pedestrian event detection method based on video of the present invention, compared with prior art, can detect pedestrian targets all in range of video, not by environmental restraint, can detect real-time video, and detection time short, be easy to realize, accuracy is higher, be well suited for real-time detect lines people event, have broad application prospects.

Accompanying drawing explanation

Fig. 1 is the first frame video image.

Fig. 2 is connected component labeling schematic diagram, and in figure, a is first piece of connected domain, and b is second piece of connected domain.

Fig. 3 is the first two field picture binaryzation mark result images, and in figure, white portion is the binaryzation target-marking of present frame.

Fig. 4 is the second frame video image, and the white crosses point in figure is the target signature point asked for.

Fig. 5 is slip scan schematic diagram, and solid-line rectangle frame is to be searched piece, and dotted line is held and represented region of search, and its central point A is the position of previous frame unique point, and solid-line rectangle frame is step-length according to a pixel, according to from left to right, and slip scan from top to bottom.

Fig. 6 is the 60th frame binaryzation signature, and the white lines in figure are pursuit path line.

Fig. 7 is the actual motion geometric locus figure that in Fig. 6, pursuit path is corresponding, and the horizontal ordinate in figure is the time, and the unit interval is 0.04s, and ordinate is actual range, and unit is cm.

Fig. 8 is pedestrian's event detection outcome.

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

Embodiment

The present embodiment provides a kind of road pedestrian event detection method based on gripper path analysis, is chosen, target trajectory matched jamming and ask for pedestrian target speed with least square fitting thus determine whether pedestrian's event by block-based binarization segmentation, block-based connected component labeling, unique point.It should be noted that, image handled in procedure of the present invention be in video along positive seasonal effect in time series first two field picture, the second two field picture, the 3rd two field picture ..., M (M is positive integer) two field picture.

If the size of each frame video image is W*H, the size of each piece is w*h, and wherein W is the pixel in each frame video video image level direction, and H is the pixel of each frame video image vertical direction, w is the width in each piece of region, and h is the height in each piece of region.

It should be noted that the mapping table in the present embodiment adopts the video camera geometric calibration method described in patent of invention " a kind of video camera geometric calibration method under linear model " (open (bulletin) number: CN102222332A) to obtain.

The method of the present embodiment specifically adopts following steps to realize:

Step one, sets up the mapping relations of image pixel to road surface actual range, i.e. mapping table;

Step 2, first two field picture is all divided into multiple pieces of regions with background image under identical block coordinate system, and the size of background is W*H, and the block size of division is w*h, the block areal T divided is T=(W/w) * (H/h), all pixel value B in i-th block _irepresent, total number N=w*h, the B of pixel in i-th block _imiddle preservation w*h pixel value, with the lower left corner of i-th block for initial point sets up right angle two-dimensional coordinate system Y, the pixel value B that in i-th block, (m, n) puts _i(m, n) represents, wherein:

W is the pixel in background level direction;

H is the pixel of background vertical direction;

W is the pixel of the Width of i-th block;

H is the pixel of the short transverse of i-th block;

i＝1,2,3...T；

M represents the horizontal ordinate of arbitrary pixel under coordinate system Y in i-th block, m=0,1,2...w-1;

N represents the ordinate of arbitrary pixel under coordinate system Y in i-th block, n=0,1,2...h-1;

Step 3, to each piece in the first two field picture, finds the background block identical with this block position in background image, and calculates the absolute value sum of the gray scale difference value of each same pixel position between this block its corresponding background block,

When the absolute value sum of gained is greater than the threshold value A of setting, then this block is object block, is 255 by the gray-scale value assignment of pixels all in this object block,

When the absolute value sum of gained is less than or equal to the threshold value A of setting, then this block is background block, is 0 by the gray-scale value assignment of pixels all in background block, wherein:

The span of described threshold value A is the area of (10 ~ 20) × block;

Finally the background in the first two field picture and target are separated, obtain the binary image of the first two field picture;

Step 4, to the binary image obtained according to from left to right, order from top to bottom scans in units of block, and adjacent object block is labeled as same target, calculate height and the width of each target-marking simultaneously, when the value of height/width is within the scope of threshold value B, rim detection is carried out to this binary image, find best angle point, namely when laterally detect data with longitudinally detect data be greater than threshold value T simultaneously time, the value of described threshold value T is 180, just judges that this point is as best angle point, wherein:

The scope of described threshold value B is 2 ~ 10;

Step 5, to select in these angle points laterally to detect data and longitudinally detect data and minimum point as clarification of objective point, create an object construction body simultaneously, record these clarification of objective point position and matched jamming counter informations, matched jamming counter R first time is initialized as zero, and unique point coordinate is (x ₁, y ₁), record the image of a block size centered by this unique point as template, the pixel value two-dimensional array B of all pixels in this template simultaneously _t[N] represents, in this block, the pixel value of pixel (m, n) is B _t(m, n);

Step 6, in the second two field picture, and with the image block B of the first frame recording _t[N] is template, with the characteristic point position (x of this template ₁, y ₁) centered by, choose the square area of 4 block sizes as region of search at current frame image, according to from left to right, method from top to bottom, searches for one by one, and the number of to be searched piece is designated as N, uses B _j[N] represents any one the search block in current search region, obtains both definitely value difference and SAD,

Wherein:

$\mathrm{SAD}=\underset{m=0}{\overset{w-1}{\mathrm{\Σ}}}\underset{n=0}{\overset{h-1}{\mathrm{\Σ}}}|B_j(m,n)-B_t(m,n)|$

j＝1,2,3....N

Minimum as matching criterior using absolute length chang, choose N number of absolute value difference and in minimum block be match block, be designated as B _s[N], namely have found matching characteristic point in the current frame, records the position (x of new matching characteristic point simultaneously ₂, y ₂), and in order to the block image B centered by new unique point _s[N] is as new template B _t[N], matched jamming counter R adds 1 simultaneously;

Step 7, from the 3rd two field picture to M two field picture, M be more than or equal to 60 positive integer, repeat step 6, when R equals threshold value C, perform step 8;

Step 8, when matched jamming counter R equals threshold value C, all unique point (x ₁, y ₁) ... (x ₆₀, y ₆₀) form the pursuit path of target, by the mapping table set up in finding step one, obtain the actual range that pursuit path is corresponding, consecutive point be spaced apart every frame time, so can obtain the array of one group of time and actual range, utilize least square fitting can try to achieve the speed of pursuit path, when this speed meets the scope of threshold value D, namely determine that this target is for pedestrian, wherein:

The value of described threshold value C is 60;

The scope of described threshold value D is (0.3 ~ 2.0) m/s;

Below provide specific embodiments of the invention, it should be noted that the present invention is not limited to following specific embodiment, all equivalents done on technical scheme basis all fall into protection scope of the present invention.

Embodiment:

In embodiment, in processing procedure, the sample frequency of video is that 25 frames are per second, every two field picture size is 720 × 288, the size in every block region is 8 × 6, two field picture is divided into 90 × 48 block regions, and target area binarization segmentation threshold value A is 576, the scope of threshold value B is 2 ~ 10, the value of threshold value T be 180, threshold value C value be 60, threshold value D scope be (0.3 ~ 2.0) m/s, as shown in Figures 1 to 8, defer to said method to process to the 60 two field picture the first frame successively.

As can be seen from Figure 6 in figure white wire be pedestrian from the first frame to the movement locus of the 60 frame, the lower end of this track enters scene, the characteristic point position found for pedestrian first time, and topmost point is that the 60th frame mates the unique point found.

Fig. 7 is the actual range curve map that Fig. 6 pursuit path is corresponding, adopt least square method to this section of track fitting, the actual motion speed 1.46m/s of pedestrian can be tried to achieve, as shown in Figure 8, this movement velocity is in the scope of threshold value D, and therefore testing result is pedestrian's event of this road.

Claims (1)

1. based on a road pedestrian event detection method for gripper path analysis, it is characterized in that, the method is carried out according to following steps:

Step one, sets up the mapping relations of image pixel to road surface actual range, i.e. mapping table;

Step 2, first two field picture is all divided into multiple pieces of regions with background image under identical block coordinate system, and the size of background is W*H, and the block size of division is w*h, the block areal T divided is T=(W/w) * (H/h), all pixel value B in i-th block _irepresent, total number N=w*h, the B of pixel in i-th block _imiddle preservation w*h pixel value, with the lower left corner of i-th block for initial point sets up right angle two-dimensional coordinate system Y, the pixel value B that in i-th block, (m, n) puts _i(m, n) represents, wherein:

W is the pixel in background level direction;

H is the pixel of background vertical direction;

W is the pixel of the Width of i-th block;

H is the pixel of the short transverse of i-th block;

i＝1,2,3...T；

M represents the horizontal ordinate of arbitrary pixel under coordinate system Y in i-th block, m=0,1,2...w-1;

N represents the ordinate of arbitrary pixel under coordinate system Y in i-th block, n=0,1,2...h-1;

Step 3, to each piece in the first two field picture, finds the background block identical with this block position in background image, and calculates the absolute value sum of the gray scale difference value of each same pixel position between this block its corresponding background block,

When the absolute value sum of gained is greater than the threshold value A of setting, then this block is object block, is 255 by the gray-scale value assignment of pixels all in this object block;

When the absolute value sum of gained is less than or equal to the threshold value A of setting, then this block is background block, is 0 by the gray-scale value assignment of pixels all in background block, wherein:

The span of described threshold value A is the area of (10 ~ 20) × block;

Finally the background in the first two field picture and target are separated, obtain the binary image of the first two field picture;

Step 4, to the binary image obtained according to from left to right, order from top to bottom scans in units of block, and adjacent object block is labeled as same target, calculates height and the width of each target-marking simultaneously, when the value of height/width is within the scope of threshold value B, rim detection is carried out to this binary image, finds best angle point, namely when laterally detect data with longitudinally detect data be greater than threshold value T simultaneously time, judge that this point is as best angle point, wherein:

The scope of described threshold value B is 2 ~ 10;

The value of described threshold value T is 180;

Step 5, to select in these angle points laterally to detect data and longitudinally detect data and minimum point as clarification of objective point, create an object construction body simultaneously, record these clarification of objective point position and matched jamming counter informations, matched jamming counter R first time is initialized as zero, and unique point coordinate is (x ₁, y ₁), record the image of a block size centered by this unique point as template, the pixel value two-dimensional array B of all pixels in this template simultaneously _t[N] represents, in this block, the pixel value of pixel (m, n) is B _t(m, n);

Step 6, in the second two field picture, with the image block B of the first frame recording _t[N] is template, with the characteristic point position (x of this template ₁, y ₁) centered by, the square area choosing 4 block sizes in current frame image is as region of search, and according to from left to right, method from top to bottom, searches for one by one, and the number of to be searched piece is designated as N, uses B _j[N] represents any one the search block in current search region, obtains both definitely value difference and SAD,

Wherein:

$\mathrm{SAD}=\underset{m=0}{\overset{w-1}{\mathrm{\Σ}}}\underset{n=0}{\overset{h-1}{\mathrm{\Σ}}}|B_j(m,n)-B_t(m,n)|$

j＝1,2,3....N

Minimum as matching criterior using absolute length chang, choose N number of absolute value difference and in minimum block be match block, be designated as B _s[N], namely have found matching characteristic point in the current frame, records the position (x of new matching characteristic point simultaneously ₂, y ₂), and in order to the block image B centered by new unique point _s[N] is as new template B _t[N], matched jamming counter R adds 1 simultaneously;

Step 7, from the 3rd two field picture to M two field picture, M be greater than 60 positive integer, repeat step 6, when R equals threshold value C, perform step 8;

Step 8, when matched jamming counter R equals threshold value C, all unique point (x ₁, y ₁) ... (x ₆₀, y ₆₀) form the pursuit path of target, by the mapping table set up in finding step one, obtain the actual range that pursuit path is corresponding, consecutive point be spaced apart every frame time, so can obtain the array of one group of time and actual range, utilize least square fitting can try to achieve the speed of pursuit path, when this speed meets the scope of threshold value D, namely determine that this target is for pedestrian, wherein:

The value of described threshold value C is 60;

The scope of described threshold value D is (0.3 ~ 2.0) m/s.