CN104778453A - Night pedestrian detection method based on statistical features of infrared pedestrian brightness - Google Patents
Night pedestrian detection method based on statistical features of infrared pedestrian brightness Download PDFInfo
- Publication number
- CN104778453A CN104778453A CN201510154382.8A CN201510154382A CN104778453A CN 104778453 A CN104778453 A CN 104778453A CN 201510154382 A CN201510154382 A CN 201510154382A CN 104778453 A CN104778453 A CN 104778453A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- frame
- infrared
- image
- infrared image
- 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
Abstract
The invention discloses a night pedestrian detection method based on statistical features of infrared pedestrian brightness. According to the method, gray average information of parts of one pedestrian and a negative sample in a sample database is firstly subjected to statistical processing, a mapping interval boundary is determined by the aid of the information, and a brightness histogram feature for distinguishing vote interval division is constructed; then, a histogram feature in the gradient direction is calculated, and the two features are combined to form a final feature descriptor; secondarily, model training is performed by the aid of Adaboost in combination of a decision tree, and pedestrian determination and positioning are performed with a sliding window scanning method; finally, when a classifier obtains lower degree of confidence through classification judgment on a certain detection frame, a brightness interval template is adopted for secondary judgment, so that night pedestrian detection is realized. Pedestrians in the night environment are effectively detected, and the method has the characteristics of high detection rate and adaptability.
Description
Technical field
The present invention relates to the pedestrian detection method of Vehicular video image, a kind of especially pedestrian detection method at night based on infrared pedestrian's brightness statistics feature.
Background technology
Pedestrian detection technology is an important application of computer vision, has very high practical value in daily life with in producing.The pedestrian detection of view-based access control model is exactly judge from the picture inputted or sequence of frames of video the particular location that pedestrian occurs according to certain image processing techniques.Intelligent vehicle DAS (Driver Assistant System) can improve the security of vehicle drive, thus reduces the generation of traffic hazard, and pedestrian detection technology is then one of core technology in intelligent DAS (Driver Assistant System).
View-based access control model night, pedestrian detection technology mainly adopted is the technology such as visible images, infrared image.In night situation, because the conditions such as illumination are undesirable, the imaging effect of visible light camera is poor, affects the effect of pedestrian detection.Thermal camera catches infrared ray information Perception object by passive infrared technology, and the object of different temperatures presents different brightness in the picture.In road scene infrared image, pedestrian is heat more more than background radiation generally, and the pedestrian in infrared image is generally more bright than background, and dark by light at night, greasy weather sight line is unclear etc. affects, there is good Infravision, have stronger adaptive faculty to different photoenvironments.Therefore the pedestrian detection technology of infrared image is the effective workaround realizing pedestrian detection at night.
The infrared pedestrian detection technology of current majority adopts the method based on machine learning, as patent of invention " pedestrian detection method of feature based combination and device " (CN103632170A) utilizes HOG feature and LBP structural feature union feature, utilize support vector machine (SVM) to carry out sorter training as learning algorithm and realize pedestrian detection.The LBP feature that said method relates to has stronger descriptive power to image texture, and the textural characteristics of infrared image not obvious, therefore the effect of LBP feature application in infrared pedestrian detection is general.Patent of invention " a kind of pedestrian detection method based on infrared image " (CN103902976A) merges and brightness histogram feature interpretation pedestrian in HOG feature base, utilizes SVM to carry out sorter training and realizes pedestrian detection at night.The method, except the profile information utilizing pedestrian, is also extracted the monochrome information of infrared pedestrian, therefore, has better infrared pedestrian detection effect.
Said method does not consider the Luminance Distribution statistical nature of infrared pedestrian.The present invention is in HOG feature base, merge the Luminance Distribution statistical nature of infrared pedestrian, build infrared pedestrian's feature that a kind of descriptive power is stronger, utilize Adaboost as learning algorithm, pedestrian pedestrian's framework that design pedestrian detection sorter and brightness section template judge again.Method of the present invention has better pedestrian's feature interpretation ability and Detection results in infrared pedestrian detection.
Summary of the invention
The object of the invention is for the deficiencies in the prior art, a kind of pedestrian detection method at night based on infrared pedestrian's brightness statistics feature is provided.The present invention utilizes profile information and the monochrome information of infrared pedestrian, first by Sample Storehouse pedestrian's parts (comprise head, above the waist, the lower part of the body) and negative sample image statistics gray average information, determine each ballot mapping range of pedestrian's Expressive Features according to the distribution characteristics of gray average information, construct the brightness histogram feature (DBHOI) that is distinguished ballot interval division; Then compute gradient direction histogram (HOG) feature, and these two features are carried out combine the final pedestrian's feature descriptor of formation; Secondly, utilize Adaboost to carry out model training in conjunction with the method for decision tree, carry out pedestrian's judgement and location by sliding window scanning method; Finally, when sorter to certain pedestrian's frame classification judgement obtain comparatively low confidence time, adopt brightness section template again to judge, thus realize night infrared pedestrian detection.The method has high, the adaptable feature of verification and measurement ratio.
The technical scheme steps that the present invention solves the employing of its technical matters is as follows:
The positive and negative sample data collection of step 1, structure infrared image;
Step 2, infrared image DBHOI latent structure based on brightness statistics feature;
Step 3, infrared image HOG latent structure based on profile information;
Step 4, Adaboost is utilized to carry out sorter training;
Step 5, detection window judge and location;
Step 6, the major punishment of pedestrian's frame are determined.
The positive and negative sample data collection of the structure infrared image described in step 1 is specific as follows:
Positive sample data collection construction method is as follows: adopt minimum rectangle window, extracts the pedestrian's sample in infrared image; Suppose that the height of pedestrian is h, wide is w, makes w/h=0.41; Extract the positive sample N of pedestrian altogether
1;
Negative sample data set construction method is as follows: random at N
0open not comprise in the infrared image of pedestrian and extract N altogether
0× 10 negative samples, namely randomly draw 10 negative samples often opening in infrared image;
Zooming to wide by all positive negative samples is 64 pixels, and height is the sample image of 128 pixels.
The infrared image DBHOI latent structure based on brightness statistics feature described in step 2 is specific as follows:
2-1. establishes sample image to be of a size of w ' × h ', then sample image is divided into the identical topography of many parts of sizes; The size of each topography is 8 × 8, is divided into the individual topography in equal (w '/8) × (h '/8), and these topographies are designated as cell by sample image;
2-2. is according to pedestrian's parts determination mapping ruler;
2-2-1. first according to N
1individual positive sample intercepts each parts of pedestrian, each parts of pedestrian comprise head, above the waist, the lower part of the body;
2-2-2. intercepts N at random in negative sample
1individual background image;
2-2-3. calculates the gray average of head image, above the waist image, lower part of the body image and background image respectively;
2-2-4. sets the gray average of four kinds of images as G
1, G
2, G
3, G
4, and size increases successively; , determine three mapped boundaries according to four gray averages, be respectively t
1=(G
1+ G
2)/2, t
2=(G
2+ G
3)/2, t
3=(G
3+ G
4)/2, are divided into intensity value ranges between four gray areas according to mapped boundaries, are respectively [0, t
1), [t
1, t
2), [t
2, t
3), [t
3, 255];
2-2-5., according to this mapping ruler and with the gray-scale value of pixel in infrared image for weights, builds brightness histogram in each cell, obtains a four-dimensional proper vector;
2 × 2 cell that 2-3. is adjacent carry out normalization in block;
With a cell for step-length, adjacent four cell are carried out normalization in block by L1-Sqrt method with order from top to bottom, from left to right by the individual cell in (w '/8) × (h '/8) in sample image;
Wherein L1-Sqrt method is as follows:
v is brightness histogram vector, and ε is a small value, and value is 0.001;
The feature of all pieces is carried out series connection and is obtained DBHOI feature by 2-4.;
Described DBHOI is the brightness histogram of different interval size.
The infrared image HOG latent structure in profile information described in step 3 is specific as follows:
Classical sobel operator is utilized to calculate the gradient component G of each pixel level and vertical direction
x(i, j), G
y(i, j), the gradient magnitude G (i, j) and direction D (i, j) that then calculate correspondence are as follows:
Described in step 4 utilize Adaboost carry out sorter training specific as follows:
All positive negative samples are zoomed to identical yardstick by 4-1., and then for each sample image, extract DBHOI proper vector and HOG proper vector, and the label marking positive sample is 1, the label of negative sample is-1;
4-2. is by N
1individual positive sample and N
0the DBHOI-HOG proper vector that × 10 negative samples are corresponding and sample label are input to Adaboost learning algorithm and train, and obtain a sorter having a series of Weak Classifier and form.
Detection window described in step 5 judges and is positioned such that:
5-1., according to the size determination zoom factor of the yardstick of infrared image to be detected and detection window, determines the convergent-divergent number of plies of infrared image to be detected; Then the sorter obtained by step 4 is that 4 pixel sizes successively scan by step-length; If the original size of infrared image to be detected is W
i× H
i, wherein W
irepresent the width of infrared image to be detected, H
irepresent infrared image height to be detected, detection window size is W
d× H
d, wherein W
drepresent the width of detection window, H
drepresent detection window height, use s
srepresent zoom factor; Then original zoom factor is s
s=1, termination zoom factor is s
s=min{W
i/ W
d, H
i/ H
d; Window scanning is slided to the infrared image to be detected under each yardstick, utilizes and train the pedestrian's sorter obtained to carry out detection window judgement; If this detection window is pedestrian's frame, then record position and the degree of confidence of this detection window, this record is expressed as { posX, posY, width, height, score}, wherein posX, posY is the upper left angle point of pedestrian's frame, width, height are width and the height of pedestrian's frame, and score is degree of confidence;
Described degree of confidence is the error rate of all two-stage decision leaf node records;
5-2. is by after carrying out multiple dimensioned sliding window Scanning Detction to infrared image to be detected, same pedestrian is detected in the infrared image to be detected of different scale, adopts the pedestrian frame result of non-maxima suppression method to multiple different scale to merge;
The standard that described maximum value suppresses is the degree of confidence score of each pedestrian's frame;
All pedestrian's frames carry out being arranged in array A [n] by degree of confidence by 5-3. from low to high, then take out as the maximum detection window information A [n] of previous belief from this array A [n];
5-4. judges this pedestrian's frame A [n] and the relation of follow-up pedestrian's frame A [n-1], if the registration α of two pedestrian's frames is greater than 0.5, then thinks same pedestrian's frame, otherwise using pedestrian's frame A [n-1] as working as the maximum pedestrian's frame of previous belief;
Described registration
wherein area (B
n) be the area of n-th pedestrian's frame, area (B
n∩ B
n-1) be the common factor of n-th pedestrian's frame area and (n-1)th pedestrian's frame area;
5-5. repeats step 5-4, until judged all pedestrian's frames.
Pedestrian's frame major punishment described in step 6 is surely as follows:
6-1. is by after step 5, and each detection window being judged as pedestrian can have a confidence value score; Have employed brightness section template to judge again pedestrian's frame; Work as score
iduring≤τ, then carry out pedestrian's frame and judge, wherein the value of threshold tau is determined by statistical; If score
i> τ, then only carry out pedestrian detection with sorter, and suppose the degree of confidence obtaining N number of pedestrian's frame, so τ meets formula
Described score
ifor the confidence value of i-th in the array A [n] of step 5-3 definition;
6-2. according to step 2 and infrared pedestrian's imaging characteristics, [0, t
1), [t
1, t
2), [t
2, t
3), [t
3, 255] these 4 interval corresponding backgrounds respectively, above the waist, the lower part of the body, head; If the wide height of pedestrian is w × h, if pedestrian's frame correctly judges pedestrian position, so the height h of the pedestrian's frame being judged as pedestrian is divided into 4 deciles from top to bottom, then must exists at first 1/4 equally divided position and belong to [t
3, 255] and the monochrome information of scope, and remember that gray-scale value is Σ p in the total number of the pixel in this interval
1, must [t be there is in second 1/4 equally divided position
1, t
2) monochrome information of scope, and remember that gray-scale value is Σ p in the total number of the pixel in this interval
2, and must [t be there is in latter two 1/4 equally divided position
2, t
3) monochrome information in scope, and remember that gray-scale value is Σ p in the total number of the pixel in this interval
3; Following proof rule is set: Σ p
1/ (w × h)>=1/16, Σ p
2/ (w × h)>=1/8, Σ p
3/ (w × h)>=1/16; If Σ is p
1, Σ p
2, Σ p
3meet this three conditions, so this pedestrian's frame is judged as pedestrian's frame again simultaneously, otherwise this pedestrian's frame is judged to be non-pedestrian frame.Beneficial effect of the present invention:
The present invention is directed to the feature of pedestrian in infrared image, construct brightness descriptor---the DBHOI that has more ability to express, and it is combined with HOG feature.DBHOI descriptor by the Luminance Distribution of each parts of pedestrian and background in statistics training sample, and is encoded to this distributed intelligence when structural attitude, makes this feature descriptor more can depict pedestrian and the difference of background in monochrome information.The make of this feature descriptor improves the descriptive power of feature, thus makes final sorter have stronger classification capacity.
When use sorter pedestrian's frame is judged after obtain comparatively low confidence time, then carry out pedestrian's frame according to brightness section template and judge again.The method greatly reduces the error rate caused because of sorter erroneous judgement, improves accuracy of detection.
Accompanying drawing explanation
Fig. 1 is overall flow figure of the present invention.
Fig. 2 is that DBHOI feature descriptor extracts process flow diagram.
Fig. 3 is that the Detection results of pedestrian detection model of the present invention is shown.
Embodiment
Below in conjunction with accompanying drawing, specific embodiment of the invention scheme is described in further detail.
As shown in Figure 1, Figure 2 and Figure 3, a kind of pedestrian detection method at night based on infrared pedestrian's brightness statistics feature, it specifically comprises the steps:
The positive and negative sample data collection of step 1, structure infrared image.
The present invention mainly for the pedestrian detection under automobile scenarios, so gather the infrared image of scene mainly under road scene of image.
Positive sample data collection construction method is as follows: adopt minimum rectangle window, that is just can surround the rectangle frame of pedestrian with one, extracts the pedestrian's sample in infrared image; Suppose that the height of pedestrian is h, wide is w, makes w/h=0.41; Extract the positive sample N of pedestrian altogether
1;
Negative sample data set construction method is as follows: random at N
0open not comprise in the infrared image of pedestrian and extract N altogether
0× 10 negative samples, namely randomly draw 10 negative samples often opening in infrared image;
Zooming to wide by all positive negative samples is 64 pixels, and height is the sample image of 128 pixels.
The definition of described minimum rectangle window is if author P.Doll á r is as described in document " PedestrianDetection:An Evaluation of the State of the Art ".
Step 2: based on the infrared image DBHOI latent structure of brightness statistics feature.
Described DBHOI is the brightness histogram of different interval size, and full name is Different BinsHistogram of Intensity.
As shown in Figure 2, the present invention is by encoding to the monochrome information in infrared image, and monochrome information is become descriptor that one has stronger resolution characteristic, its concrete make is as follows.
2-1. supposes that sample image is of a size of w ' × h ', then sample image is divided into the identical topography of many parts of sizes; The size of each topography is 8 × 8, is divided into the individual topography in equal (w '/8) × (h '/8), and these topographies are designated as cell by sample image.
2-2. is according to pedestrian's parts determination mapping ruler.First according to N
1individual positive sample intercepts each parts of pedestrian, each parts of pedestrian comprise head, above the waist, the lower part of the body; N is intercepted at random again in negative sample
1individual background image, then calculates the gray average of head image, above the waist image, lower part of the body image and background image respectively.If the gray average of four kinds of images is G
1, G
2, G
3, G
4, and size increases successively; , determine three mapped boundaries according to four gray averages, be respectively t
1=(G
1+ G
2)/2, t
2=(G
2+ G
3)/2, t
3=(G
3+ G
4)/2, are divided into intensity value ranges between four gray areas according to mapped boundaries, are respectively [0, t
1), [t
1, t
2), [t
2, t
3), [t
3, 255].According to this mapping ruler, and be weights with gray-scale value, in each cell, build brightness histogram, obtain a four-dimensional proper vector.
2 × 2 cell that 2-3. is adjacent carry out normalization in block.
Adjacent four cell, with a cell for step-length, are carried out block in normalization with order from top to bottom, from left to right by the individual cell in (w '/8) × (h '/8) by individual cell in sample image.Method for normalizing is L1-Sqrt:
wherein v is brightness histogram vector, and ε is a small value, and in the present invention, value is 0.001.
The feature of all pieces is carried out series connection and is obtained DBHOI feature by 2-4..The DBHOI descriptor built by such mode has stronger descriptive power.
Step 3, infrared image HOG latent structure based on profile information.
Classical sobel operator is utilized to calculate the gradient component G of each pixel level and vertical direction
x(i, j), G
y(i, j), the gradient magnitude G (i, j) and direction D (i, j) that then calculate correspondence are as follows:
The structure of HOG feature adopts the building method of golden allusion quotation, such as the sample image of 64 × 128 sizes is divided into several cell, and the size of each cell is 8 × 8.For each cell region, gradient direction is divided into 9 impartial deciles, and gradient magnitude is built gradient orientation histogram as weights.Adjacent 4 cell are combined into block, and L1-sqrt method is normalized, the histogram of gradients of finally connecting between different masses, forms final HOG descriptor.
Step 4: utilize Adaboost to carry out sorter training.
All positive negative samples are zoomed to identical yardstick, as 64 × 128.Then for each sample image, extract DBHOI proper vector and HOG proper vector, and the label marking positive sample is 1, the label of negative sample is-1.
By N
1individual positive sample and N
0the DBHOI-HOG proper vector that × 10 negative samples are corresponding and sample label are input to Adaboost learning algorithm and train, and obtain a sorter having a series of Weak Classifier and form.Wherein, Weak Classifier is two-stage decision tree-model.
Step 5: pedestrian's frame judges and location.
As shown in Figure 3, because the size of pedestrian's sorter is determined, but the size of pedestrian in infrared image is different.Such as when pedestrian's distance video camera is very near, its size is by large for the size than sorter.Therefore in order to detect the pedestrian under different scale, needing to carry out convergent-divergent to image, then carrying out drawing window scanning for the image under each yardstick.Same pedestrian is likely classified device and is judged to be pedestrian under different graphical rules, therefore needs the fusion carrying out pedestrian's frame.
5-1., according to the size determination zoom factor of the yardstick of infrared image to be detected and detection window, determines the convergent-divergent number of plies of infrared image to be detected; Then the sorter obtained by step 4 is that 4 pixel sizes successively scan by step-length.If the original size of infrared image to be detected is W
i× H
i, wherein W
irepresent the width of infrared image to be detected, H
irepresent infrared image height to be detected, detection window size is W
d× H
d, wherein W
drepresent the width of detection window, H
drepresent detection window height, use s
srepresent zoom factor.Then original zoom factor is s
s=1, termination zoom factor is s
s=min{W
i/ W
d, H
i/ H
d.Window scanning is slided to the infrared image to be detected under each yardstick, utilizes and train the pedestrian's sorter obtained to carry out window judgement.If this detection window is pedestrian's frame, then record position and the degree of confidence of this detection window, this record is expressed as { posX, posY, width, height, score}, wherein posX, posY is the upper left angle point of pedestrian's frame, width, height are width and the height of pedestrian's frame, and score is degree of confidence.
Described degree of confidence is the error rate of all two-stage decision leaf node records.
5-2. is by after carrying out multiple dimensioned sliding window Scanning Detction to infrared image to be detected, same pedestrian is detected in the infrared image to be detected of different scale, in order to the pedestrian's frame making system export the corresponding actual pedestrian position of a most probable, the pedestrian frame result of non-maxima suppression method to multiple different scale is adopted to merge.The standard that described maximum value suppresses is the degree of confidence score of each pedestrian's frame.
All pedestrian's frames carry out being arranged in array A [n] by degree of confidence by 5-3. from low to high, then take out as the maximum detection window information A [n] of previous belief from this array A [n];
5-4. judges this pedestrian's frame A [n] and the relation of follow-up pedestrian's frame A [n-1], if the registration α of two pedestrian's frames is greater than 0.5, then thinks same pedestrian's frame, otherwise using pedestrian's frame A [n-1] as working as the maximum pedestrian's frame of previous belief.
Described registration
wherein area (B
n) be the area of n-th pedestrian's frame, area (B
n∩ B
n-1) be the common factor of n-th pedestrian's frame area and (n-1)th pedestrian's frame area.
5-5. repeats step 5-4, until judged all pedestrian's frames.
Step 6, the major punishment of pedestrian's frame are determined.
6-1. is by after step 5, and each detection window being judged as pedestrian can have a confidence value score.If the judgement of sorter to a detection window is clearer and more definite, then corresponding score value is also larger; On the contrary, if the judgement of sorter to detection window is more indefinite, then can obtain a very low score value.Therefore, when a region is pedestrian's frame by the judgement of mistake, the degree of confidence score that this pedestrian's frame obtains is lower.This because sorter is when degree of confidence is inadequate in order to reduce, the mistake made judges, present invention employs brightness section template decision technology again.Work as score
iduring≤τ, then carry out pedestrian's frame and judge, wherein the value of threshold tau is determined by statistical.If score
i> τ, then only carry out pedestrian detection with sorter, and suppose the degree of confidence (default value N is 1000) obtaining N number of pedestrian's frame, so τ meets formula
Described score
ifor the confidence value of i-th in the array A [n] of step 5-3 definition.
6-2. according to step 2 and infrared pedestrian's imaging characteristics, [0, t
1), [t
1, t
2), [t
2, t
3), [t
3, 255] these 4 interval corresponding backgrounds respectively, above the waist, the lower part of the body, head.If the wide height of pedestrian is w × h, if pedestrian's frame correctly judges pedestrian position, so the height h of the pedestrian's frame being judged as pedestrian is divided into 4 deciles from top to bottom, then must exists at first 1/4 equally divided position and belong to [t
3, 255] and the monochrome information of scope, and remember that gray-scale value is Σ p in the total number of the pixel in this interval
1, must [t be there is in second 1/4 equally divided position
1, t
2) monochrome information of scope, and remember that gray-scale value is Σ p in the total number of the pixel in this interval
2, and must [t be there is in latter two 1/4 equally divided position
2, t
3) monochrome information in scope, and remember that gray-scale value is Σ p in the total number of the pixel in this interval
3.Following proof rule is set: Σ p
1/ (w × h)>=1/16, Σ p
2/ (w × h)>=1/8, Σ p
3/ (w × h)>=1/16.If Σ is p
1, Σ p
2, Σ p
3meet this three conditions, so this pedestrian's frame is judged as pedestrian's frame simultaneously, otherwise this pedestrian's frame is judged to be non-pedestrian frame.
Claims (7)
1., based on a pedestrian detection method at night for infrared pedestrian's brightness statistics feature, it is characterized in that comprising the steps:
The positive and negative sample data collection of step 1, structure infrared image;
Step 2, infrared image DBHOI latent structure based on brightness statistics feature;
Step 3, infrared image HOG latent structure based on profile information;
Step 4, Adaboost is utilized to carry out sorter training;
Step 5, detection window judge and location;
Step 6, the major punishment of pedestrian's frame are determined.
2. a kind of pedestrian detection method at night based on infrared pedestrian's brightness statistics feature according to claim 1, is characterized in that the positive and negative sample data collection of the structure infrared image described in step 1 is specific as follows:
Positive sample data collection construction method is as follows: adopt minimum rectangle window, extracts the pedestrian's sample in infrared image; Suppose that the height of pedestrian is h, wide is w, makes w/h=0.41; Extract the positive sample N of pedestrian altogether
1;
Negative sample data set construction method is as follows: random at N
0open not comprise in the infrared image of pedestrian and extract N altogether
0× 10 negative samples, namely randomly draw 10 negative samples often opening in infrared image;
Zooming to wide by all positive negative samples is 64 pixels, and height is the sample image of 128 pixels.
3. a kind of pedestrian detection method at night based on infrared pedestrian's brightness statistics feature according to claim 1, is characterized in that the infrared image DBHOI latent structure based on brightness statistics feature described in step 2 is specific as follows:
2-1. establishes sample image to be of a size of w ' × h ', then sample image is divided into the identical topography of many parts of sizes; The size of each topography is 8 × 8, is divided into the individual topography in equal (w '/8) × (h '/8), and these topographies are designated as cell by sample image;
2-2. is according to pedestrian's parts determination mapping ruler;
2-2-1. first according to N
1individual positive sample intercepts each parts of pedestrian, each parts of pedestrian comprise head, above the waist, the lower part of the body;
2-2-2. intercepts N at random in negative sample
1individual background image;
2-2-3. calculates the gray average of head image, above the waist image, lower part of the body image and background image respectively;
2-2-4. sets the gray average of four kinds of images as G
1, G
2, G
3, G
4, and size increases successively; , determine three mapped boundaries according to four gray averages, be respectively t
1=(G
1+ G
2)/2, t
2=(G
2+ G
3)/2, t
3=(G
3+ G
4)/2, are divided into intensity value ranges between four gray areas according to mapped boundaries, are respectively [0, t
1), [t
1, t
2), [t
2, t
3), [t
3, 255];
2-2-5., according to this mapping ruler and with the gray-scale value of pixel in infrared image for weights, builds brightness histogram in each cell, obtains a four-dimensional proper vector;
2 × 2 cell that 2-3. is adjacent carry out normalization in block;
With a cell for step-length, adjacent four cell are carried out normalization in block by L1-Sqrt method with order from top to bottom, from left to right by the individual cell in (w '/8) × (h '/8) in sample image;
Wherein L1-Sqrt method is as follows:
v is brightness histogram vector, and ε is a small value, and value is 0.001;
The feature of all pieces is carried out series connection and is obtained DBHOI feature by 2-4.;
Described DBHOI is the brightness histogram of different interval size.
4. a kind of pedestrian detection method at night based on infrared pedestrian's brightness statistics feature according to claim 1, is characterized in that the infrared image HOG latent structure in profile information described in step 3 is specific as follows:
Classical sobel operator is utilized to calculate the gradient component G of each pixel level and vertical direction
x(i, j), G
y(i, j), the gradient magnitude G (i, j) and direction D (i, j) that then calculate correspondence are as follows:
5. a kind of pedestrian detection method at night based on infrared pedestrian's brightness statistics feature according to claim 1, it is characterized in that described in step 4 utilize Adaboost carry out sorter training specific as follows:
All positive negative samples are zoomed to identical yardstick by 4-1., and then for each sample image, extract DBHOI proper vector and HOG proper vector, and the label marking positive sample is 1, the label of negative sample is-1;
4-2. is by N
1individual positive sample and N
0the DBHOI-HOG proper vector that × 10 negative samples are corresponding and sample label are input to Adaboost learning algorithm and train, and obtain a sorter having a series of Weak Classifier and form.
6. a kind of pedestrian detection method at night based on infrared pedestrian's brightness statistics feature according to claim 1, is characterized in that the detection window described in step 5 judges and is positioned such that:
5-1., according to the size determination zoom factor of the yardstick of infrared image to be detected and detection window, determines the convergent-divergent number of plies of infrared image to be detected; Then the sorter obtained by step 4 is that 4 pixel sizes successively scan by step-length; If the original size of infrared image to be detected is W
i× H
i, wherein W
irepresent the width of infrared image to be detected, H
irepresent infrared image height to be detected, detection window size is W
d× H
d, wherein W
drepresent the width of detection window, H
drepresent detection window height, use s
srepresent zoom factor; Then original zoom factor is s
s=1, termination zoom factor is s
s=min{W
i/ W
d, H
i/ H
d; Window scanning is slided to the infrared image to be detected under each yardstick, utilizes and train the pedestrian's sorter obtained to carry out detection window judgement; If this detection window is pedestrian's frame, then record position and the degree of confidence of this detection window, this record is expressed as { posX, posY, width, height, score}, wherein posX, posY is the upper left angle point of pedestrian's frame, width, height are width and the height of pedestrian's frame, and score is degree of confidence;
Described degree of confidence is the error rate of all two-stage decision leaf node records;
5-2. is by after carrying out multiple dimensioned sliding window Scanning Detction to infrared image to be detected, same pedestrian is detected in the infrared image to be detected of different scale, adopts the pedestrian frame result of non-maxima suppression method to multiple different scale to merge;
The standard that described maximum value suppresses is the degree of confidence score of each pedestrian's frame;
All pedestrian's frames carry out being arranged in array A [n] by degree of confidence by 5-3. from low to high, then take out as the maximum detection window information A [n] of previous belief from this array A [n];
5-4. judges this pedestrian's frame A [n] and the relation of follow-up pedestrian's frame A [n-1], if the registration α of two pedestrian's frames is greater than 0.5, then thinks same pedestrian's frame, otherwise using pedestrian's frame A [n-1] as working as the maximum pedestrian's frame of previous belief;
Described registration
wherein area (B
n) be the area of n-th pedestrian's frame, area (B
n∩ B
n-1) be the common factor of n-th pedestrian's frame area and (n-1)th pedestrian's frame area;
5-5. repeats step 5-4, until judged all pedestrian's frames.
7. a kind of pedestrian detection method at night based on infrared pedestrian's brightness statistics feature according to claim 1, is characterized in that the pedestrian's frame major punishment described in step 6 is surely as follows:
6-1. is by after step 5, and each detection window being judged as pedestrian can have a confidence value score; Have employed brightness section template to judge again pedestrian's frame; Work as score
iduring≤τ, then carry out pedestrian's frame and judge, wherein the value of threshold tau is determined by statistical; If score
i> τ, then only carry out pedestrian detection with sorter, and suppose the degree of confidence obtaining N number of pedestrian's frame, so τ meets formula
Described score
ifor the confidence value of i-th in the array A [n] of step 5-3 definition;
6-2. according to step 2 and infrared pedestrian's imaging characteristics, [0, t
1), [t
1, t
2), [t
2, t
3), [t
3, 255] these 4 interval corresponding backgrounds respectively, above the waist, the lower part of the body, head; If the wide height of pedestrian is w × h, if pedestrian's frame correctly judges pedestrian position, so the height h of the pedestrian's frame being judged as pedestrian is divided into 4 deciles from top to bottom, then must exists at first 1/4 equally divided position and belong to [t
3, 255] and the monochrome information of scope, and remember that gray-scale value is Σ p in the total number of the pixel in this interval
1, must [t be there is in second 1/4 equally divided position
1, t
2) monochrome information of scope, and remember that gray-scale value is Σ p in the total number of the pixel in this interval
2, and must [t be there is in latter two 1/4 equally divided position
2, t
3) monochrome information in scope, and remember that gray-scale value is Σ p in the total number of the pixel in this interval
3; Following proof rule is set: Σ p
1/ (w × h)>=1/16, Σ p
2/ (w × h)>=1/8, Σ p
3/ (w × h)>=1/16; If Σ is p
1, Σ p
2, Σ p
3meet this three conditions, so this pedestrian's frame is judged as pedestrian's frame again simultaneously, otherwise this pedestrian's frame is judged to be non-pedestrian frame.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510154382.8A CN104778453B (en) | 2015-04-02 | 2015-04-02 | A kind of night pedestrian detection method based on infrared pedestrian's brightness statistics feature |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510154382.8A CN104778453B (en) | 2015-04-02 | 2015-04-02 | A kind of night pedestrian detection method based on infrared pedestrian's brightness statistics feature |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104778453A true CN104778453A (en) | 2015-07-15 |
| CN104778453B CN104778453B (en) | 2017-12-22 |
Family
ID=53619906
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510154382.8A Active CN104778453B (en) | 2015-04-02 | 2015-04-02 | A kind of night pedestrian detection method based on infrared pedestrian's brightness statistics feature |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104778453B (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105139427A (en) * | 2015-09-10 | 2015-12-09 | 华南理工大学 | Part segmentation method suitable for re-identification of pedestrian video |
| CN105139869A (en) * | 2015-07-27 | 2015-12-09 | 安徽清新互联信息科技有限公司 | Baby crying detection method based on interval difference features |
| CN105279754A (en) * | 2015-09-10 | 2016-01-27 | 华南理工大学 | Part segmentation method suitable for bicycle video detection |
| CN105389546A (en) * | 2015-10-22 | 2016-03-09 | 四川膨旭科技有限公司 | System for identifying person at night during vehicle driving process |
| CN105426852A (en) * | 2015-11-23 | 2016-03-23 | 天津津航技术物理研究所 | Method for identifying pedestrians by vehicle-mounted monocular long-wave infrared camera |
| CN105631410A (en) * | 2015-12-18 | 2016-06-01 | 华南理工大学 | Classroom detection method based on intelligent video processing technology |
| CN105913003A (en) * | 2016-04-07 | 2016-08-31 | 国家电网公司 | Multi-characteristic multi-model pedestrian detection method |
| CN106372624A (en) * | 2016-10-15 | 2017-02-01 | 杭州艾米机器人有限公司 | Human face recognition method and human face recognition system |
| CN106600635A (en) * | 2016-11-03 | 2017-04-26 | 上海机电工程研究所 | Infrared target radiation characteristic simulation model checking verifying method based on small subsamples |
| CN106919930A (en) * | 2017-03-10 | 2017-07-04 | 成都智锗科技有限公司 | A kind of low resolution infrared image parahypnosis situation determination methods |
| CN108734178A (en) * | 2018-05-18 | 2018-11-02 | 电子科技大学 | A kind of HOG feature extracting methods of rule-basedization template |
| CN109034174A (en) * | 2017-06-08 | 2018-12-18 | 北京君正集成电路股份有限公司 | A kind of cascade classifier training method and device |
| CN111126444A (en) * | 2019-11-28 | 2020-05-08 | 天津津航技术物理研究所 | Classifier integration method |
| CN111382718A (en) * | 2020-03-17 | 2020-07-07 | 电子科技大学中山学院 | Night pedestrian detection system and pedestrian detection method based on system |
| CN111508003A (en) * | 2020-04-20 | 2020-08-07 | 北京理工大学 | Infrared small target detection tracking and identification method |
| CN113095120A (en) * | 2020-01-09 | 2021-07-09 | 北京君正集成电路股份有限公司 | System for realizing reduction of upper human body detection false alarm |
| CN113313078A (en) * | 2021-07-02 | 2021-08-27 | 昆明理工大学 | Lightweight night infrared image pedestrian detection method and system based on model optimization |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120154580A1 (en) * | 2010-12-20 | 2012-06-21 | Huang tai-hui | Moving object detection method and image processing system for moving object detection |
| CN102609682A (en) * | 2012-01-13 | 2012-07-25 | 北京邮电大学 | Feedback pedestrian detection method for region of interest |
| CN103632170A (en) * | 2012-08-20 | 2014-03-12 | 深圳市汉华安道科技有限责任公司 | Pedestrian detection method and device based on characteristic combination |
| CN103886308A (en) * | 2014-04-15 | 2014-06-25 | 中南大学 | Pedestrian detection method through soft cascade classifiers according to polymerization channel characteristics |
| CN103902976A (en) * | 2014-03-31 | 2014-07-02 | 浙江大学 | Pedestrian detection method based on infrared image |
-
2015
- 2015-04-02 CN CN201510154382.8A patent/CN104778453B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120154580A1 (en) * | 2010-12-20 | 2012-06-21 | Huang tai-hui | Moving object detection method and image processing system for moving object detection |
| CN102609682A (en) * | 2012-01-13 | 2012-07-25 | 北京邮电大学 | Feedback pedestrian detection method for region of interest |
| CN103632170A (en) * | 2012-08-20 | 2014-03-12 | 深圳市汉华安道科技有限责任公司 | Pedestrian detection method and device based on characteristic combination |
| CN103902976A (en) * | 2014-03-31 | 2014-07-02 | 浙江大学 | Pedestrian detection method based on infrared image |
| CN103886308A (en) * | 2014-04-15 | 2014-06-25 | 中南大学 | Pedestrian detection method through soft cascade classifiers according to polymerization channel characteristics |
Non-Patent Citations (4)
| Title |
|---|
| AMNON SHASHUA 等: "Pedestrian Detection for Driving Assistance Systems: Single-frame Classification and System Level Performance", 《2004 IEEE INTELLIGENT VEHICLES SYMPOSIUM》 * |
| E. BINELLI 等: "A modular tracking system for far infrared pedestrian recognition", 《2005 PROCEEDINGS IEEE INTELLIGENT VEHICLES SYMPOSIUM》 * |
| 朱聪聪: "夜晚环境下的行人检测技术研究", 《中国优秀硕士学位论文全文数据库·信息科技辑》 * |
| 赵君钦 等: "红外图像中人体目标检测技术研究", 《现代电子技术》 * |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105139869A (en) * | 2015-07-27 | 2015-12-09 | 安徽清新互联信息科技有限公司 | Baby crying detection method based on interval difference features |
| CN105279754B (en) * | 2015-09-10 | 2018-06-22 | 华南理工大学 | A kind of component dividing method suitable for bicycle video detection |
| CN105279754A (en) * | 2015-09-10 | 2016-01-27 | 华南理工大学 | Part segmentation method suitable for bicycle video detection |
| CN105139427A (en) * | 2015-09-10 | 2015-12-09 | 华南理工大学 | Part segmentation method suitable for re-identification of pedestrian video |
| CN105139427B (en) * | 2015-09-10 | 2018-06-22 | 华南理工大学 | A kind of component dividing method identified again suitable for pedestrian's video |
| CN105389546A (en) * | 2015-10-22 | 2016-03-09 | 四川膨旭科技有限公司 | System for identifying person at night during vehicle driving process |
| CN105426852A (en) * | 2015-11-23 | 2016-03-23 | 天津津航技术物理研究所 | Method for identifying pedestrians by vehicle-mounted monocular long-wave infrared camera |
| CN105426852B (en) * | 2015-11-23 | 2019-01-08 | 天津津航技术物理研究所 | Vehicle-mounted monocular LONG WAVE INFRARED camera pedestrian recognition method |
| CN105631410A (en) * | 2015-12-18 | 2016-06-01 | 华南理工大学 | Classroom detection method based on intelligent video processing technology |
| CN105631410B (en) * | 2015-12-18 | 2019-04-09 | 华南理工大学 | A kind of classroom detection method based on intelligent video processing technique |
| CN105913003A (en) * | 2016-04-07 | 2016-08-31 | 国家电网公司 | Multi-characteristic multi-model pedestrian detection method |
| CN105913003B (en) * | 2016-04-07 | 2019-06-07 | 国家电网公司 | A kind of pedestrian detection method of multiple features multi-model |
| CN106372624A (en) * | 2016-10-15 | 2017-02-01 | 杭州艾米机器人有限公司 | Human face recognition method and human face recognition system |
| CN106600635A (en) * | 2016-11-03 | 2017-04-26 | 上海机电工程研究所 | Infrared target radiation characteristic simulation model checking verifying method based on small subsamples |
| CN106919930A (en) * | 2017-03-10 | 2017-07-04 | 成都智锗科技有限公司 | A kind of low resolution infrared image parahypnosis situation determination methods |
| CN109034174B (en) * | 2017-06-08 | 2021-07-09 | 北京君正集成电路股份有限公司 | Cascade classifier training method and device |
| CN109034174A (en) * | 2017-06-08 | 2018-12-18 | 北京君正集成电路股份有限公司 | A kind of cascade classifier training method and device |
| CN108734178A (en) * | 2018-05-18 | 2018-11-02 | 电子科技大学 | A kind of HOG feature extracting methods of rule-basedization template |
| CN111126444A (en) * | 2019-11-28 | 2020-05-08 | 天津津航技术物理研究所 | Classifier integration method |
| CN113095120A (en) * | 2020-01-09 | 2021-07-09 | 北京君正集成电路股份有限公司 | System for realizing reduction of upper human body detection false alarm |
| CN111382718A (en) * | 2020-03-17 | 2020-07-07 | 电子科技大学中山学院 | Night pedestrian detection system and pedestrian detection method based on system |
| CN111382718B (en) * | 2020-03-17 | 2023-08-11 | 电子科技大学中山学院 | Night pedestrian detection system and pedestrian detection method based on same |
| CN111508003A (en) * | 2020-04-20 | 2020-08-07 | 北京理工大学 | Infrared small target detection tracking and identification method |
| CN113313078A (en) * | 2021-07-02 | 2021-08-27 | 昆明理工大学 | Lightweight night infrared image pedestrian detection method and system based on model optimization |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104778453B (en) | 2017-12-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104778453A (en) | Night pedestrian detection method based on statistical features of infrared pedestrian brightness | |
| CN103902976B (en) | A kind of pedestrian detection method based on infrared image | |
| CN105046196B (en) | Front truck information of vehicles structuring output method based on concatenated convolutional neutral net | |
| CN103886308B (en) | A kind of pedestrian detection method of use converging channels feature and soft cascade grader | |
| CN103761531B (en) | The sparse coding license plate character recognition method of Shape-based interpolation contour feature | |
| CN109190444B (en) | Method for realizing video-based toll lane vehicle feature recognition system | |
| CN107122776A (en) | A kind of road traffic sign detection and recognition methods based on convolutional neural networks | |
| CN103971097B (en) | Vehicle license plate recognition method and system based on multiscale stroke models | |
| CN106127137A (en) | A kind of target detection recognizer based on 3D trajectory analysis | |
| CN105447503B (en) | Pedestrian detection method based on rarefaction representation LBP and HOG fusion | |
| CN106503748A (en) | A kind of based on S SIFT features and the vehicle targets of SVM training aids | |
| CN102214291A (en) | Method for quickly and accurately detecting and tracking human face based on video sequence | |
| CN105809121A (en) | Multi-characteristic synergic traffic sign detection and identification method | |
| CN111611905A (en) | Visible light and infrared fused target identification method | |
| CN106886778B (en) | License plate character segmentation and recognition method in monitoring scene | |
| CN104021375A (en) | Model identification method based on machine learning | |
| CN105760858A (en) | Pedestrian detection method and apparatus based on Haar-like intermediate layer filtering features | |
| CN104598885A (en) | Method for detecting and locating text sign in street view image | |
| CN111401188B (en) | Traffic police gesture recognition method based on human body key point characteristics | |
| CN106529461A (en) | Vehicle model identifying algorithm based on integral characteristic channel and SVM training device | |
| CN106372658A (en) | Vehicle classifier training method | |
| CN108256462A (en) | A kind of demographic method in market monitor video | |
| CN109886086A (en) | Pedestrian detection method based on HOG feature and Linear SVM cascade classifier | |
| CN104217206A (en) | Real-time attendance counting method based on high-definition videos | |
| CN104463104A (en) | Fast detecting method and device for static vehicle target |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| EXSB | Decision made by sipo to initiate substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20150715 Assignee: HANGZHOU ZHISHU TECHNOLOGY CO.,LTD. Assignor: HANGZHOU DIANZI University Contract record no.: X2022330000062 Denomination of invention: A night pedestrian detection method based on infrared pedestrian brightness statistical characteristics Granted publication date: 20171222 License type: Common License Record date: 20220331 |
|
| EE01 | Entry into force of recordation of patent licensing contract |