CN102930287A - Overlook-based detection and counting system and method for pedestrians - Google Patents
Overlook-based detection and counting system and method for pedestrians Download PDFInfo
- Publication number
- CN102930287A CN102930287A CN2012103649630A CN201210364963A CN102930287A CN 102930287 A CN102930287 A CN 102930287A CN 2012103649630 A CN2012103649630 A CN 2012103649630A CN 201210364963 A CN201210364963 A CN 201210364963A CN 102930287 A CN102930287 A CN 102930287A
- Authority
- CN
- China
- Prior art keywords
- sub
- block
- image
- gradient
- pixel
- 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.)
- Granted
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses an overlook-based detection and counting system and an overlook-based detection and counting method for pedestrians. The system comprise a camera unit, a detection unit, a training unit and a judging and counting unit, wherein the camera unit is arranged at the upper end of the top of a gateway and is used for shooting the pedestrians who pass through the gateway; and the detection unit comprises a square target frame setting module, a sliding detection window (a square frame) image gradient acquisition module, an image any pixel point gradient direction value acquisition module, a subblock pixel gradient direction counting module, a subblock uniformization module and a gradient direction information in-series superposition module. According to the system and the method, the problem of occlusion between targets is solved, so that the accuracy of detection can be improved.
Description
Technical field
The invention belongs to technical field of image processing, relate to a kind of pedestrian detection number system, relate in particular to a kind of detection number system for overlooking the pedestrian; Simultaneously, the invention still further relates to a kind of detection method of counting for overlooking the pedestrian.
Background technology
Nowadays, the at any time progress of science and technology and the requirement of social development need to arrange the pedestrian detection number system in a lot of places.The number system of traditional sense mainly is to utilize sensor to carry out, and degree of accuracy has much room for improvement.Simultaneously, process the progressively development of counting along with image, begin to occur utilizing video camera and image processing techniques to calculate the method for pedestrian's number.
The existing method of utilizing image processing techniques to detect pedestrian's number is to look squarely the pedestrian by video camera to detect, and the pedestrian who detects is trunk and the four limbs of upright hinge, and the pedestrian of back is easily blocked by the pedestrian of front, and accuracy of detection is not high.
In view of this, nowadays in the urgent need to designing a kind of new pedestrian detection number system, in order to improve the degree of accuracy that detects.
Summary of the invention
Technical matters to be solved by this invention is: a kind of detection number system for overlooking the pedestrian is provided, owing to substantially do not have the problem of blocking between the target, thus can improve the degree of accuracy of detection.
In addition, the present invention also provides a kind of detection method of counting for overlooking the pedestrian, does not substantially have the problem of blocking between the target, thereby can improve the degree of accuracy of detection.
For solving the problems of the technologies described above, the present invention adopts following technical scheme:
A kind of detection number system for overlooking the pedestrian, described system comprises:
Image unit is arranged on the top of gateway, and the pedestrian in order to shooting is passed through uses the camera better effects if with wide-angle lens;
Detecting unit comprises:
-square target frame setting module, in order to set a square target frame, it is square box, make square box confine the number of people in the zone of overhead view image, the size of square box is determined according to the actual conditions that detect target, change with the camera setting height(from bottom) is different, the rule of delineation is: the number of people of overlooking is positioned in the middle of the window, leave a determining deviation all around, the slip detection window is exactly this above-mentioned square box, detection window is divided into 3 * 3 totally 9 square sub-blocks, has half area overlapping between sub-block and the sub-block; If establish detection window length and wide be respectively w, then the length of side of square sub-block is w/2;
-slip detection window (being square box) image gradient acquisition module is in order to gray level image data and template [1,0,1] and [1,0,1] with square-shaped frame
TDo convolution, obtain any pixel of image at x direction gradient value d
x(x, y) and y direction gradient value d
y(x, y);
Any pixel gradient direction value acquisition module of-image, in order to obtain the gradient direction value θ (x, y) of any pixel (x, y) of image:
-sub-block pixel gradient directional statistics module, in order to add up the gradient direction of each sub-block pixel: the gradient direction that calculates all pixels in the sub-block drops on respectively the accumulated value in 9 intervals of-180 °～180 ° of scopes, and each interval is the scope at 40 ° of angles; The gradient direction of pixel is cumulative is with weighting coefficient, and namely the gradient direction value of each pixel will multiply by a coefficient, and this coefficient equals the absolute value sum of this pixel x direction and y direction gradient; Each sub-block obtains a row or column vector according to oneself statistical value in the gradient direction in 9 intervals; Each component that is each sub-block vector is exactly the statistical value that the gradient direction of this sub-block pixel drops on respectively 9 intervals;
-sub-block normalized module is in order to each sub-block of normalized; If v represents not normalized vector, || v|| represent the vector the single order norm, e is a little constant, then normalized formula be v=v/ (|| v||+e);
-gradient direction message linkage laminating module in order to the gradient direction information of 9 sub-blocks of a square detection block the inside successively overlapped in series, forms one 81 vector of tieing up; The Feature Conversion that is each moving window (being square box) becomes one 81 vector about the gradient direction statistics of tieing up.
As a preferred embodiment of the present invention, described system further comprises:
Training unit, in order to gather from the environment of reality, to extract the samples pictures that comprises in a large number the sample of target and do not comprise target, the sample that comprises target is positive sample, the sample that does not comprise target is negative sample; Utilize described detecting unit that all samples pictures are extracted proper vector; The size of samples pictures must be with the slip detection window big or small the same; Then, adopt support vector machine to make sorter, with the proper vector of positive and negative samples it is trained, obtain the parameter of sorter, namely obtain the discrimination model of sorter;
Data acquisition and detecting unit, the picture in order to the picture that utilizes the new collection of described detecting unit sliding window swept-volume newly gathers the scene detects forms with setting step-length interlacing or every being listed as one by one scanned picture with sliding;
Judge and counting unit, in order to extracting the proper vector of the picture region that detects behind the sliding window, and judge in the middle of being input to sorter, judge whether to contain target; If so, then count, and continue scanning by step-length; If not, then continue by step scan.
As a preferred embodiment of the present invention, described training unit also comprises the greyscale image transitions module, in order to just to convert first coloured image to gray level image.
As a preferred embodiment of the present invention, described data acquisition and detecting unit detect forms take step-length as 5 pixel interlacing or every being listed as one by one scanned picture with sliding.
As a preferred embodiment of the present invention, described data acquisition and detecting unit only detect 2～3 frame pictures each second, and according to actual needs each detection only detects the zone of picture 1/3～1/2.
A kind of above-mentioned method of counting for the detection number system of overlooking the pedestrian, described method comprises the steps:
Shooting step: by the pedestrian of image unit shooting by setting regions;
Detecting step: comprising:
-square target frame is set step: set a square-shaped frame by square target frame setting module, make square-shaped frame confine the number of people in the zone of overhead view image, the size of square-shaped frame is determined according to the actual conditions that detect target, change with the camera setting height(from bottom) is different, the rule of delineation is: the number of people of overlooking is positioned in the middle of the window, leave a determining deviation all around, the slip detection window is exactly this above-mentioned square-shaped frame, detection window is divided into 3 * 3 totally 9 little square sub-blocks, has half area overlapping between piece and the piece; Be w and h if establish the length of side of detection window, the length of side of then little square sub-block is w/2;
-slip detection window image gradient obtaining step: by gray level image data and template [1,0,1] and [1,0,1] of slip detection window image gradient acquisition module with square-shaped frame
TDo convolution, obtain any pixel of image at x direction gradient value d
x(x, y) and y direction gradient value d
y(x, y);
Any pixel gradient direction value obtaining step of-image: the gradient direction value θ (x, y) that obtains any pixel (x, y) of image by any pixel gradient direction value acquisition module of image:
-sub-block pixel gradient directional statistics step: add up the gradient direction of each sub-block pixel by sub-block pixel gradient directional statistics module: the gradient direction that calculates all pixels in the sub-block drops on respectively the accumulated value in 9 intervals of-180 °～180 ° of scopes, and each interval is the scope at 40 ° of angles; The gradient direction of pixel is cumulative is with weighting coefficient, and namely the gradient direction value of each pixel will multiply by a coefficient, and this coefficient equals the absolute value sum of this pixel x direction and y direction gradient; Each sub-block obtains a row or column vector according to from the gradient direction statistical value in 9 intervals; Each component that is each sub-block vector is exactly the statistical value that the gradient direction of this sub-block pixel drops on respectively 9 intervals;
-sub-block normalized step: by each sub-block of sub-block normalized module normalized; If v represents not normalized vector, || v|| represent the vector the single order norm, e is a little constant, then normalized formula be v=v/ (|| v||+e);
-gradient direction message linkage stack step: with the gradient direction information of 9 sub-blocks of a square-shaped frame the inside successively overlapped in series, forms one 81 vector of tieing up by gradient direction message linkage laminating module; The Feature Conversion that is each sliding window becomes one 81 vector about the gradient direction statistics of tieing up.
As a preferred embodiment of the present invention, described method further comprises:
Training step: gather from the environment of reality, extract the samples pictures that comprises in a large number the sample of target and do not comprise target, the sample that comprises target is positive sample, and the sample that does not comprise target is negative sample; Utilize described detecting unit that all samples pictures are extracted proper vector; The size of samples pictures must be with the slip detection window big or small the same; Then, adopt support vector machine to make sorter, with the proper vector of positive and negative samples it is trained, obtain the parameter of sorter, namely obtain the discrimination model of sorter;
Data acquisition and detecting step: utilize the picture of the new collection of described detecting unit sliding window swept-volume to the new picture that gathers in scene, detect forms with setting step-length interlacing or every being listed as one by one scanned picture with sliding;
Judge and counting step: extract the proper vector of the picture region that detects behind the sliding window, and judge in the middle of being input to sorter, judge whether to contain target; If so, then count, and continue scanning by step-length; If not, then continue by step scan.
As a preferred embodiment of the present invention, described training step also comprises the greyscale image transitions step, just converts first coloured image to gray level image.
As a preferred embodiment of the present invention, in described data acquisition and the detecting step, detect forms take step-length as 3～5 pixel interlacing or every being listed as one by one scanned picture with sliding.
As a preferred embodiment of the present invention, in described data acquisition and the detecting step, only detect 2～3 frame pictures each second, according to actual needs each detection only detects the zone of picture 1/3～1/2.
Beneficial effect of the present invention is: the present invention propose for detection number system and the method for overlooking the pedestrian, substantially do not have the problem of blocking between the target, thereby can improve the degree of accuracy of detection.
The feature of overlooking pedestrian head field color homogeneous that the present invention is directed to and the external appearance characteristic of edge contour design: head zone color homogeneous, mean that Grad is low, and the Grad of head edge has then reflected the variation of target wheel profile.Take the gradient direction distribution of regional area as the basis, embody its importance by the number of times that repeats to add up.Such as the zone of core, comprise the edge of the number of people, the distribution characteristics of its gradient direction repeats to have 4 times at most; The feature at four angles only occurs 1 time; Remainder occurs 2 times.That is to say, the gradient direction statistical nature of middle sub-image is crucial.Such design is to have given prominence to the target signature that detects, the impact of the peripheral background that weakened again.
Description of drawings
Fig. 1 is slip detection window inner structure and feature schematic diagram.
Fig. 2 is the process flow diagram of the inventive method detecting step.
Fig. 3 is the process flow diagram of the inventive method.
Embodiment
Describe the preferred embodiments of the present invention in detail below in conjunction with accompanying drawing.
Embodiment one
The present invention has disclosed a kind of detection number system for overlooking the pedestrian, and the present invention utilizes monocular cam to overlook under the state, to the detecting and counting of pedestrian, is mainly used in the statistics to various passenger flows.
Camera is installed in the upper end, top of public passage gateway, and camera is taken the pedestrian who passes through downwards, and the largest benefit of putting like this camera is exactly substantially not have the problem of blocking between the target.Generally apart from ground 3～10 meters of cameras are pedestrian detection of a kind of closer distance.Compare with the pedestrian detection of looking squarely of traditional sense, although detect congruence, detect clarification of objective obviously different, general pedestrian is trunk and the four limbs of upright hinge, and the pedestrian in overlooking is the agglomerate that moves one by one.
Passenger flow detection method in this paper, identify target with head feature, confine the number of people in the zone of overhead view image with a square-shaped frame, the size of square-shaped frame is determined according to the actual conditions that detect target (can change with the camera setting height(from bottom) is different), the rule of delineation is: the number of people of overlooking is positioned in the middle of the window, leave a determining deviation all around, as shown in Figure 1, the slip detection window is exactly this above-mentioned square-shaped frame, detection window is divided into 3 * 3 totally 9 little square sub-blocks, has half area overlapping between piece and the piece.If establishing the length of side of detection window is w, the length of side of then little square sub-block is w/2.The view field of overlooking the number of people should just comprise the area of that little square block of (or substantially comprising) middle.
Described system comprises: image unit, detecting unit, training unit, judgement and counting unit.
Image unit is arranged at the top upper end of gateway, the pedestrian who passes through in order to shooting; Image unit can be video camera.
Detecting unit comprises: square target frame setting module, slip detection window image gradient acquisition module, any pixel gradient direction value acquisition module of image, sub-block pixel gradient directional statistics module, sub-block normalized module, gradient direction message linkage laminating module.
The square-shaped frame setting module is in order to set a square-shaped frame, make square-shaped frame confine the number of people in the zone of overhead view image, the size of square-shaped frame is determined according to the actual conditions that detect target, change with the camera setting height(from bottom) is different, the rule of delineation is: the number of people of overlooking is positioned in the middle of the window, leaves a determining deviation all around, and the slip detection window is exactly this above-mentioned square-shaped frame, detection window is divided into 3 * 3 totally 9 little square sub-blocks, has half area overlapping between sub-block and the sub-block; If establish detection window length and wide be respectively w, the length of side of then little square sub-block is w/2.
Slip detection window image gradient acquisition module is in order to gray level image data and template [1,0,1] and [1,0,1] with square-shaped frame
TDo convolution, obtain any pixel of image at x direction gradient value d
x(x, y) and y direction gradient value d
y(x, y).
Any pixel gradient direction value acquisition module of image is in order to obtain the gradient direction value θ (x, y) of any pixel (x, y) of image:
Sub-block pixel gradient directional statistics module is in order to add up the gradient direction of each sub-block pixel: the gradient direction that calculates all pixels in the sub-block drops on respectively the accumulated value in 9 intervals of-180 °～180 ° of scopes, and each interval is the scope at 40 ° of angles; The gradient direction of pixel is cumulative is with weighting coefficient, and namely the gradient direction value of each pixel will multiply by a coefficient, and this coefficient equals the absolute value sum of this pixel x direction and y direction gradient; Each sub-block obtains a row or column vector according to the statistical value of self 9 interval gradient direction; Each component that is the corresponding vector of each sub-block is exactly the statistical value that this sub-block gradient direction drops on respectively 9 intervals.
Sub-block normalized module is in order to each sub-block of normalized; If v represents not normalized vector, || v|| represent the vector the single order norm, e is a little constant, then normalized formula be v=v/ (|| v||+e).
Gradient direction message linkage laminating module forms one 81 vector of tieing up in order to the gradient direction information of 9 sub-blocks of square-shaped frame the inside successively overlapped in series; The Feature Conversion that is each sliding window becomes one 81 vector about the gradient direction statistics of tieing up.
The samples pictures that training unit comprises the sample of target and do not comprise target in order to gather, to extract a large amount of (crossing a thousand sheets) from the environment of reality, the sample that comprises target is positive sample, the sample that does not comprise target is negative sample; Utilize described detecting unit that all samples pictures are extracted proper vector; The size of samples pictures must be with the slip detection window big or small the same; Then, adopt support vector machine to make sorter, with the proper vector of positive and negative samples it is trained, obtain the parameter of sorter, namely obtain the discrimination model of sorter.What native system was processed is gray level image, is that coloured image just converts gray level image to first if gather, and the image size is 320 * 240 pixels or less.
Data acquisition and detecting unit to the new picture that gathers in scene, detect forms take setting step-length (such as step-length as 3～5 pixels) interlacing or every being listed as one by one scanned picture in order to the new picture that gathers of sliding window swept-volume with sliding.For reducing calculated amount, only detect 2～3 frame pictures each second, according to actual needs each detection only detects the zone of picture 1/3～1/2.
Judge and counting unit in order to extracting the proper vector of the picture region that detects behind the sliding window, and judge in the middle of being input to sorter, judge whether to contain target; If so, then count, and continue scanning by step-length; If not, then continue by step scan.
More than introduced and the present invention is directed to the detection number system of overlooking the pedestrian, the present invention also discloses a kind of above-mentioned method of counting for the detection number system of overlooking the pedestrian when disclosing said system; See also Fig. 2, described method comprises:
Shooting step: by the pedestrian of image unit shooting by setting regions.
Detecting step: comprising:
-square-shaped frame is set step: set a square-shaped frame by the square-shaped frame setting module, make square-shaped frame confine the number of people in the zone of overhead view image, the size of square-shaped frame is determined according to the actual conditions that detect target, change with the camera setting height(from bottom) is different, the rule of delineation is: the number of people of overlooking is positioned in the middle of the window, leave a determining deviation all around, the slip detection window is exactly this above-mentioned square-shaped frame, detection window is divided into 3 * 3 totally 9 little square block, has half area overlapping between piece and the piece; If establish detection window length and wide be respectively w, the length of side of then little square block is w/2;
-slip detection window image gradient obtaining step: by gray level image data and template [1,0,1] and [1,0,1] of slip detection window image gradient acquisition module with square-shaped frame
TDo convolution, obtain any pixel of image at x direction gradient value d
x(x, y) and y direction gradient value d
y(x, y);
Any pixel gradient direction value obtaining step of-image: the gradient direction value θ (x, y) that obtains any pixel (x, y) of image by any pixel gradient direction value acquisition module of image:
-sub-block pixel gradient directional statistics step: add up the gradient direction of each sub-block pixel by sub-block pixel gradient directional statistics module: the gradient direction that calculates all pixels in the sub-block drops on respectively the accumulated value in 9 intervals of-180 °～180 ° of scopes, and each interval is the scope at 40 ° of angles; The gradient direction of pixel is cumulative is with weighting coefficient, and namely the gradient direction value of each pixel will multiply by a coefficient, and this coefficient equals the absolute value sum of this pixel x direction and y direction gradient; Each sub-block obtains a row or column vector according to the statistical value of self 9 interval gradient direction; Each component that is the corresponding vector of each sub-block is exactly the statistical value that this sub-block gradient direction drops on respectively 9 intervals;
-sub-block normalized step: by each sub-block of sub-block normalized module normalized; If v represents not normalized vector, || v|| represent the vector the single order norm, e is a little constant, then normalized formula be v=v/ (|| v||+e);
-gradient direction message linkage stack step: with the gradient direction information of 9 sub-blocks of a square-shaped frame the inside successively overlapped in series, forms one 81 vector of tieing up by gradient direction message linkage laminating module; The Feature Conversion that is each sliding window becomes one 81 vector about the gradient direction statistics of tieing up.
See also Fig. 3, whole method of the present invention comprises:
[training step] gathers from the environment of reality, extracts the samples pictures that comprises in a large number the sample of target and do not comprise target, and the sample that comprises target is positive sample, and the sample that does not comprise target is negative sample; Utilize described detecting unit that all samples pictures are extracted proper vector; The size of samples pictures must be with the slip detection window big or small the same; Then, adopt support vector machine to make sorter, with the proper vector of positive and negative samples it is trained, obtain the parameter of sorter, namely obtain the discrimination model of sorter;
[data acquisition and detecting step] utilizes the picture of the new collection of described detecting unit sliding window swept-volume to the new picture that gathers in scene, detects forms with setting step-length interlacing or every being listed as one by one scanned picture with sliding;
[judging and counting step] extracts the proper vector of the picture region that detects behind the sliding window, and judges in the middle of being input to sorter, judges whether to contain target; If so, then count, and continue scanning by step-length; If not, then continue by step scan.
In sum, the present invention propose for detection number system and the method for overlooking the pedestrian, substantially do not have the problem of blocking between the target, thereby can improve the degree of accuracy of detection.
The feature of overlooking pedestrian head field color homogeneous that the present invention is directed to and the external appearance characteristic of edge contour design: head zone color homogeneous, mean that Grad is low, and the Grad of head edge has then reflected the variation of target wheel profile.Take the gradient direction distribution of regional area as the basis, embody its importance by the number of times that repeats to add up.Such as the zone of core, comprise the edge of the number of people, the distribution characteristics of its gradient direction repeats to have 4 times at most; The feature at four angles only occurs 1 time; Remainder occurs 2 times.That is to say, the gradient direction statistical nature of middle sub-image is crucial.Such design is to have given prominence to the target signature that detects, the impact of the peripheral background that weakened again.
Here description of the invention and application is illustrative, is not to want with scope restriction of the present invention in the above-described embodiments.Here the distortion of disclosed embodiment and change is possible, and the various parts of the replacement of embodiment and equivalence are known for those those of ordinary skill in the art.Those skilled in the art are noted that in the situation that do not break away from spirit of the present invention or essential characteristic, and the present invention can be with other form, structure, layout, ratio, and realize with other assembly, material and parts.In the situation that do not break away from the scope of the invention and spirit, can carry out other distortion and change to disclosed embodiment here.
Claims (10)
1. one kind for the detection number system of overlooking the pedestrian, it is characterized in that, described system comprises:
Image unit is arranged on the top of gateway, and the pedestrian in order to downward shooting is passed through uses the camera better effects if with wide-angle lens;
Detecting unit comprises:
-square target frame setting module, in order to set a square target frame, it is square box, make square box confine the number of people in the zone of overhead view image, the size of square box is determined according to the actual conditions that detect target, change with the camera setting height(from bottom) is different, the rule of delineation is: the number of people of overlooking is positioned in the middle of the window, leave a determining deviation all around, the slip detection window is exactly this above-mentioned square box, detection window is divided into 3 * 3 totally 9 little square sub-blocks, has half area overlapping between sub-block and the sub-block; If establish detection window length and wide be respectively w, then the length of side of square sub-block is w/2;
-slip detection window image gradient acquisition module is in order to gray level image data and template [1,0,1] and [1,0,1] with square box
TDo convolution, obtain any pixel of image at x direction gradient value d
x(x, y) and y direction gradient value d
y(x, y);
Any pixel gradient direction value acquisition module of-image, in order to obtain the gradient direction value θ (x, y) of any pixel (x, y) of image:
-sub-block pixel gradient directional statistics module, in order to add up the gradient direction of each sub-block pixel: the gradient direction that calculates all pixels in the sub-block drops on respectively the accumulated value in 9 intervals of-180 °～180 ° of scopes, and each interval is the scope at 40 ° of angles; The cumulative weighting coefficient of being with of pixel gradient direction, namely the gradient direction value of each pixel will multiply by a coefficient, and this coefficient equals the absolute value sum of this pixel x direction and y direction gradient; Each sub-block drops on the statistical value in 9 intervals according to self gradient direction, obtain a row or column vector: namely each component of the corresponding vector of each sub-block is exactly the statistical value that this sub-block gradient direction drops on respectively 9 intervals;
-sub-block normalized module is in order to each sub-block of normalized; If v represents not normalized vector, || v|| represent the vector the single order norm, e is a little constant, then normalized formula be v=v/ (|| v||+e);
-gradient direction message linkage laminating module in order to the gradient direction information of 9 sub-blocks of square box the inside successively overlapped in series, forms one 81 vector of tieing up; The Feature Conversion that is each sliding window becomes one 81 vector about the gradient direction statistics of tieing up.
2. the detection number system for overlooking the pedestrian according to claim 1 is characterized in that:
Described system further comprises:
Training unit, in order to gather from the environment of reality, to extract the samples pictures that comprises in a large number the sample of target and do not comprise target, the sample that comprises target is positive sample, the sample that does not comprise target is negative sample; Utilize described detecting unit that all training sample pictures are extracted proper vector; The size of samples pictures must be with the slip detection window big or small the same; Then, adopt support vector machine to make sorter, with the proper vector of positive and negative samples it is trained, obtain the parameter of sorter, namely obtain the discrimination model of sorter;
Data acquisition and detecting unit, the picture in order to the picture that utilizes the new collection of described detecting unit sliding window swept-volume newly gathers the scene detects forms with setting step-length interlacing or every being listed as one by one scanned picture with sliding;
Judge and counting unit, in order to extracting the proper vector of the picture region that detects behind the sliding window, and judge in the middle of being input to sorter, judge whether to contain target; If so, then count, and continue scanning by step-length; If not, then continue by step scan.
3. the detection number system for overlooking the pedestrian according to claim 2 is characterized in that:
Native system only needs to process with gray level image, if collection is coloured image, described training unit also comprises the greyscale image transitions module, in order to just to convert first coloured image to gray level image.
4. the detection number system for overlooking the pedestrian according to claim 2 is characterized in that:
Described data acquisition and detecting unit detect forms take step-length as 3～5 pixel interlacing or every being listed as one by one scanned picture with sliding.
5. the detection number system for overlooking the pedestrian according to claim 4 is characterized in that:
Described data acquisition and detecting unit only detect 2～3 frame pictures each second, and according to actual needs each detection only detects the zone of picture 1/3～1/2.
6. the method for counting of the detection number system for overlooking the pedestrian claimed in claim 1 is characterized in that, described method comprises the steps:
Shooting step: by the pedestrian of image unit shooting by setting regions;
Detecting step: comprising:
-square target frame is set step: set a square target frame by square target frame setting module, make square box confine the number of people in the zone of overhead view image, the size of square box is determined according to the actual conditions that detect target, change with the camera setting height(from bottom) is different, the rule of delineation is: the number of people of overlooking is positioned in the middle of the window, leave a determining deviation all around, the slip detection window is exactly this above-mentioned square box, detection window is divided into 3 * 3 totally 9 little square sub-blocks, has half area overlapping between sub-block and the sub-block; If establish detection window length and wide be respectively w, then little square block-length of side is w/2;
-slip detection window image gradient obtaining step: by gray level image data and template [1,0,1] and [1,0,1] of slip detection window image gradient acquisition module with detection block
TDo convolution, obtain any pixel of image at x direction gradient value d
x(x, y) and y direction gradient value d
y(x, y);
Any pixel gradient direction value obtaining step of-image: the gradient direction value θ (x, y) that obtains any pixel (x, y) of image by any pixel gradient direction value acquisition module of image:
-sub-block pixel gradient directional statistics step: add up the gradient direction of each sub-block pixel by sub-block pixel gradient directional statistics module: the gradient direction that calculates all pixels in the sub-block drops on respectively the accumulated value in 9 intervals of-180 °～180 ° of scopes, and each interval is the scope at 40 ° of angles; The gradient direction of pixel is cumulative is with weighting coefficient, and namely the gradient direction value of each pixel will multiply by a coefficient, and this coefficient equals the absolute value sum of this pixel x direction and y direction gradient; Each sub-block obtains a row or column vector according to oneself statistical value in the gradient direction in 9 intervals; Each component that is the corresponding vector of each sub-block is exactly the statistical value that this sub-block gradient direction drops on respectively 9 intervals;
-sub-block normalized step: by each sub-block of sub-block normalized module normalized; If v represents not normalized vector, || v|| represent the vector the single order norm, e is a little constant, then normalized formula be v=v/ (|| v||+e);
-gradient direction message linkage stack step: with the gradient direction information of 9 sub-blocks of a square box the inside successively overlapped in series, forms one 81 vector of tieing up by gradient direction message linkage laminating module; The Feature Conversion that is each sliding window becomes one 81 vector about the gradient direction statistics of tieing up.
7. method of counting according to claim 6 is characterized in that:
Described method further comprises:
Training step: gather from the environment of reality, extract the samples pictures that comprises in a large number the sample of target and do not comprise target, the sample that comprises target is positive sample, and the sample that does not comprise target is negative sample; Utilize described detecting unit that all samples pictures are extracted proper vector; The size of samples pictures must be with the slip detection window big or small the same; Then, adopt support vector machine to make sorter, with the proper vector of positive and negative samples it is trained, obtain the parameter of sorter, namely obtain the discrimination model of sorter;
Data acquisition and detecting step: utilize the picture of the new collection of described detecting unit sliding window swept-volume to the new picture that gathers in scene, detect forms with setting step-length interlacing or every being listed as one by one scanned picture with sliding; Judge and counting step: extract the proper vector of the picture region that detects behind the sliding window, and judge in the middle of being input to sorter, judge whether to contain target; If so, then count, and continue scanning by step-length; If not, then continue by step scan.
8. method of counting according to claim 7 is characterized in that:
Described training step also comprises the greyscale image transitions step, just converts first coloured image to gray level image.
9. method of counting according to claim 7 is characterized in that:
In described data acquisition and the detecting step, detect forms take step-length as 5 pixel interlacing or every being listed as one by one scanned picture with sliding.
10. method of counting according to claim 9 is characterized in that:
In described data acquisition and the detecting step, only detect 2～3 frame pictures each second, according to actual needs each detection only detects the zone of picture 1/3～1/2.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210364963.0A CN102930287B (en) | 2012-09-26 | 2012-09-26 | A kind of detection number system and method for overlooking pedestrian |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210364963.0A CN102930287B (en) | 2012-09-26 | 2012-09-26 | A kind of detection number system and method for overlooking pedestrian |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102930287A true CN102930287A (en) | 2013-02-13 |
| CN102930287B CN102930287B (en) | 2015-09-02 |
Family
ID=47645084
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201210364963.0A Expired - Fee Related CN102930287B (en) | 2012-09-26 | 2012-09-26 | A kind of detection number system and method for overlooking pedestrian |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102930287B (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103400142A (en) * | 2013-05-14 | 2013-11-20 | 上海交通大学 | Pedestrian counting method |
| CN103473953A (en) * | 2013-08-28 | 2013-12-25 | 奇瑞汽车股份有限公司 | Pedestrian detection method and system |
| CN103559492A (en) * | 2013-11-12 | 2014-02-05 | 公安部第三研究所 | Car logo recognition device and method |
| CN103559478A (en) * | 2013-10-07 | 2014-02-05 | 唐春晖 | Passenger flow counting and event analysis method for video monitoring of pedestrians in overlooking mode |
| CN104020848A (en) * | 2014-05-15 | 2014-09-03 | 中航华东光电（上海）有限公司 | Static gesture recognizing method |
| CN104239905A (en) * | 2013-06-17 | 2014-12-24 | 上海盖普电梯有限公司 | Moving target recognition method and intelligent elevator billing system having moving target recognition function |
| CN105046316A (en) * | 2015-05-12 | 2015-11-11 | 中山大学 | Laser bidirectional pedestrian counting method based on Gaussian process regression |
| CN106951885A (en) * | 2017-04-08 | 2017-07-14 | 广西师范大学 | A kind of people flow rate statistical method based on video analysis |
| CN108388883A (en) * | 2018-03-16 | 2018-08-10 | 广西师范大学 | A kind of video demographic method based on HOG+SVM |
| CN109059863A (en) * | 2018-06-29 | 2018-12-21 | 大连民族大学 | Method of the pedestrian track point DUAL PROBLEMS OF VECTOR MAPPING to two-dimensional world coordinate system will be looked squarely |
| CN109670519A (en) * | 2017-10-13 | 2019-04-23 | 佳能株式会社 | Image processing apparatus and image processing method |
| CN109934208A (en) * | 2019-04-22 | 2019-06-25 | 江苏邦融微电子有限公司 | A kind of hardware-accelerated system and method for fingerprint recognition |
| CN110472593A (en) * | 2019-08-20 | 2019-11-19 | 重庆紫光华山智安科技有限公司 | Training image acquisition methods, model training method and relevant apparatus |
| CN111402632A (en) * | 2020-03-18 | 2020-07-10 | 五邑大学 | Risk prediction method for pedestrian movement track at intersection |
| CN111899180A (en) * | 2019-05-05 | 2020-11-06 | 上海闻通信息科技有限公司 | Image key pixel direction positioning method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1426898A2 (en) * | 2002-12-06 | 2004-06-09 | Samsung Electronics Co., Ltd. | Human detection through face detection and motion detection |
| CN101655910A (en) * | 2008-08-21 | 2010-02-24 | 索尼（中国）有限公司 | Training system, training method and detection method |
-
2012
- 2012-09-26 CN CN201210364963.0A patent/CN102930287B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1426898A2 (en) * | 2002-12-06 | 2004-06-09 | Samsung Electronics Co., Ltd. | Human detection through face detection and motion detection |
| CN101655910A (en) * | 2008-08-21 | 2010-02-24 | 索尼（中国）有限公司 | Training system, training method and detection method |
Non-Patent Citations (1)
| Title |
|---|
| 黄朝露: ""基于机器视觉的客流量统计技术研究"", 《中国优秀硕士学位论文全文数据库（电子期刊）》 * |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103400142A (en) * | 2013-05-14 | 2013-11-20 | 上海交通大学 | Pedestrian counting method |
| CN103400142B (en) * | 2013-05-14 | 2016-06-01 | 上海交通大学 | A kind of pedestrian counting method |
| CN104239905A (en) * | 2013-06-17 | 2014-12-24 | 上海盖普电梯有限公司 | Moving target recognition method and intelligent elevator billing system having moving target recognition function |
| CN103473953A (en) * | 2013-08-28 | 2013-12-25 | 奇瑞汽车股份有限公司 | Pedestrian detection method and system |
| CN103473953B (en) * | 2013-08-28 | 2015-12-09 | 奇瑞汽车股份有限公司 | A kind of pedestrian detection method and system |
| CN103559478A (en) * | 2013-10-07 | 2014-02-05 | 唐春晖 | Passenger flow counting and event analysis method for video monitoring of pedestrians in overlooking mode |
| CN103559478B (en) * | 2013-10-07 | 2018-12-04 | 唐春晖 | Overlook the passenger flow counting and affair analytical method in pedestrian's video monitoring |
| CN103559492A (en) * | 2013-11-12 | 2014-02-05 | 公安部第三研究所 | Car logo recognition device and method |
| CN104020848A (en) * | 2014-05-15 | 2014-09-03 | 中航华东光电（上海）有限公司 | Static gesture recognizing method |
| CN105046316B (en) * | 2015-05-12 | 2018-08-03 | 中山大学 | A kind of two-way pedestrian counting method of laser returned based on Gaussian process |
| CN105046316A (en) * | 2015-05-12 | 2015-11-11 | 中山大学 | Laser bidirectional pedestrian counting method based on Gaussian process regression |
| CN106951885A (en) * | 2017-04-08 | 2017-07-14 | 广西师范大学 | A kind of people flow rate statistical method based on video analysis |
| CN109670519A (en) * | 2017-10-13 | 2019-04-23 | 佳能株式会社 | Image processing apparatus and image processing method |
| CN109670519B (en) * | 2017-10-13 | 2023-09-26 | 佳能株式会社 | Image processing apparatus and image processing method |
| CN108388883A (en) * | 2018-03-16 | 2018-08-10 | 广西师范大学 | A kind of video demographic method based on HOG+SVM |
| CN109059863A (en) * | 2018-06-29 | 2018-12-21 | 大连民族大学 | Method of the pedestrian track point DUAL PROBLEMS OF VECTOR MAPPING to two-dimensional world coordinate system will be looked squarely |
| CN109059863B (en) * | 2018-06-29 | 2020-09-22 | 大连民族大学 | Method for mapping track point vector of head-up pedestrian to two-dimensional world coordinate system |
| CN109934208A (en) * | 2019-04-22 | 2019-06-25 | 江苏邦融微电子有限公司 | A kind of hardware-accelerated system and method for fingerprint recognition |
| CN111899180A (en) * | 2019-05-05 | 2020-11-06 | 上海闻通信息科技有限公司 | Image key pixel direction positioning method |
| CN111899180B (en) * | 2019-05-05 | 2023-11-17 | 上海闻通信息科技有限公司 | Image key pixel direction positioning method |
| CN110472593A (en) * | 2019-08-20 | 2019-11-19 | 重庆紫光华山智安科技有限公司 | Training image acquisition methods, model training method and relevant apparatus |
| CN111402632A (en) * | 2020-03-18 | 2020-07-10 | 五邑大学 | Risk prediction method for pedestrian movement track at intersection |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102930287B (en) | 2015-09-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102930287B (en) | A kind of detection number system and method for overlooking pedestrian | |
| CN102163284B (en) | Chinese environment-oriented complex scene text positioning method | |
| CN103886308B (en) | A kind of pedestrian detection method of use converging channels feature and soft cascade grader | |
| CN102542289B (en) | Pedestrian volume statistical method based on plurality of Gaussian counting models | |
| CN101916383B (en) | Vehicle detecting, tracking and identifying system based on multi-camera | |
| CN106127137A (en) | A kind of target detection recognizer based on 3D trajectory analysis | |
| CN102663357A (en) | Color characteristic-based detection algorithm for stall at parking lot | |
| CN104077577A (en) | Trademark detection method based on convolutional neural network | |
| CN102622584B (en) | Method for detecting mask faces in video monitor | |
| CN103268470B (en) | Object video real-time statistical method based on any scene | |
| CN106339657B (en) | Crop straw burning monitoring method based on monitor video, device | |
| CN101527043B (en) | Video picture segmentation method based on moving target outline information | |
| CN102163278B (en) | Illegal vehicle intruding detection method for bus lane | |
| CN102842045A (en) | Pedestrian detection method based on combined features | |
| CN102915433A (en) | Character combination-based license plate positioning and identifying method | |
| CN104933424A (en) | Vehicle and pedestrian monitoring method and apparatus | |
| CN103065163B (en) | A kind of fast target based on static images detects recognition system and method | |
| CN101714213A (en) | Vehicle-mounted pedestrian detection method and system | |
| CN106709438A (en) | Method for collecting statistics of number of people based on video conference | |
| CN104463232A (en) | Density crowd counting method based on HOG characteristic and color histogram characteristic | |
| CN102749034B (en) | Railway switch gap offset detection method based on image processing | |
| CN103279752A (en) | Eye locating method based on improved Adaboost algorithm and human face geometrical characteristics | |
| CN102436578B (en) | Formation method for dog face characteristic detector as well as dog face detection method and device | |
| CN101867729B (en) | Method for detecting news video formal soliloquy scene based on features of characters | |
| CN103514460A (en) | Video monitoring multi-view-angle vehicle detecting method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20150902 Termination date: 20180926 |