CN102201059A - Pedestrian detection method and device - Google Patents
Pedestrian detection method and device Download PDFInfo
- Publication number
- CN102201059A CN102201059A CN 201110132332 CN201110132332A CN102201059A CN 102201059 A CN102201059 A CN 102201059A CN 201110132332 CN201110132332 CN 201110132332 CN 201110132332 A CN201110132332 A CN 201110132332A CN 102201059 A CN102201059 A CN 102201059A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- image
- zone
- interest
- adaboost cascade
- 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.)
- Pending
Links
Images
Abstract
The invention discloses a pedestrian detection method and a pedestrian detection device. The method comprises the following steps of: collecting a pedestrian image, preprocessing the image, and extracting a region of interest from the preprocessed image; pre-positioning a pedestrian region in the region of interest by adopting an Adaboost cascade classifier; and positioning the pedestrian region in the pedestrian region pre-positioned by the Adaboost cascade classifier by adopting a support vector machine (SVM). By the technical scheme, the technical problem that the fallout ratio in the prior art is high.
Description
Technical field
The present invention relates to technical field of image processing, relate in particular to a kind of pedestrian detection method and device.
Background technology
Along with development of electronic technology, pedestrian detection technology based on image or video has replaced traditional modes such as infrared or radar gradually, possessing higher reliability, convenience and low cost based on the pedestrian detection technology of image or video more and more receives publicity, existing pedestrian detection technology based on image or video generally adopts the Adaboost cascade classifier that the pedestrian is detected, though the pedestrian detection technology that this pedestrian detection technology is more traditional has obtained good effect, there is the higher shortcoming of false drop rate.
Summary of the invention
The invention provides a kind of pedestrian detection method and device, solve the higher technical matters of false drop rate in the prior art.
For solving the problems of the technologies described above, the present invention by the following technical solutions:
A kind of pedestrian detection method comprises:
Gather pedestrian's image, from described image, extract area-of-interest;
Adopt the pedestrian zone in the described area-of-interest of Adaboost cascade classifier pre-determined bit;
Adopt support vector machine svm classifier device in the pedestrian zone of described Adaboost cascade classifier pre-determined bit, further to locate the pedestrian zone.
Described Adaboost cascade classifier comprises whole Adaboost cascade classifier of pedestrian that the whole sample of employing pedestrian becomes with the respective negative sample training and/or the local Adaboost cascade classifier of pedestrian that adopts pedestrian's fractional sample to become with the respective negative sample training; Described svm classifier device comprises whole svm classifier device of pedestrian that the whole sample of employing pedestrian becomes with the respective negative sample training and/or the local svm classifier device of pedestrian that adopts pedestrian's fractional sample to become with the respective negative sample training.
Gather pedestrian's image, the method of extracting area-of-interest from described image is specially: the pedestrian's image that collects is divided at least two image blocks, from the current image frame that is divided into two image blocks, select the image block that will detect at least according to default image block cycle detection order, and with this image block as area-of-interest.
Adopt support vector machine svm classifier device in the pedestrian zone of described Adaboost cascade classifier pre-determined bit further after the pedestrian zone, location, also comprise the step that follow the tracks of in pedestrian zone that further location is obtained, comprising:
Judge whether to navigate to the pedestrian zone, enlarge according to setting size or ratio if then will further locate the pedestrian zone that obtains;
When from the next frame image, extracting area-of-interest, with the pedestrian zone that enlarges in the next frame image institute corresponding region and the image block that will detect from the next frame picture frame, selected according to image block cycle detection order together as area-of-interest.
The mode of gathering pedestrian's image comprises: the pedestrian's image that utilizes the camera collection peripheral vehicle that is arranged at body of a motor car.
Utilize the pedestrian of the preposition camera collection vehicle front of the automobile image of leaning to one side.
A kind of pedestrian detection device comprises image capture module, image processing module, Adaboost cascade classifier and support vector machine svm classifier device, wherein,
Described image capture module is used to gather pedestrian's image;
Described image processing module is used for extracting area-of-interest from described image;
Described Adaboost cascade classifier is used for the pedestrian zone of the described area-of-interest of pre-determined bit;
Described support vector machine svm classifier device is used for further locating the pedestrian zone in the pedestrian zone of described Adaboost cascade classifier pre-determined bit.
Described Adaboost cascade classifier comprises whole Adaboost cascade classifier of pedestrian that the whole sample of employing pedestrian becomes with the respective negative sample training and/or the local Adaboost cascade classifier of pedestrian that adopts pedestrian's fractional sample to become with respective negative sample white silk; Described svm classifier device comprises whole svm classifier device of pedestrian that the whole sample of employing pedestrian becomes with the respective negative sample training and/or the local svm classifier device of pedestrian that adopts pedestrian's fractional sample to become with respective negative sample white silk.
Pedestrian's image that described image processing module specifically is used for collecting is divided at least two image blocks, from the current image frame that is divided into two image blocks, select the image block that will detect at least according to default image block cycle detection order, and with this image block as area-of-interest.
Described image capture module comprises the camera that is arranged at body of a motor car, is used to gather pedestrian's image of peripheral vehicle.
The invention provides a kind of pedestrian detection method and device, this method is at first carried out pre-determined bit by the Adaboost cascade classifier to the pedestrian zone in the area-of-interest, by support vector machine svm classifier device the pedestrian zone of pre-determined bit is further detected again, further accurately locate the pedestrian zone, the present invention is by the cooperation of Adaboost cascade classifier and support vector machine svm classifier device, than the mode that only pedestrian is detected in the prior art by the Adaboost cascade classifier, reduce false drop rate, improved detection efficiency.
Further, whole Adaboost cascade classifier of pedestrian that the whole sample of Adaboost cascade classifier employing pedestrian of the present invention becomes with the respective negative sample training and/or the local Adaboost cascade classifier of pedestrian that adopts pedestrian's fractional sample to become with respective negative sample white silk, support vector machine svm classifier device comprises whole svm classifier device of pedestrian that the whole sample of employing pedestrian becomes with the respective negative sample training and/or the local svm classifier device of pedestrian that adopts pedestrian's fractional sample to become with respective negative sample white silk, the sorter that becomes with the respective negative sample training than the whole sample of available technology adopting pedestrian, the sorter that adopts fractional sample to become with the respective negative sample training can further reduce false drop rate, promotes detection speed simultaneously.
Further, the present invention extracts area-of-interest from image method comprises: the pedestrian's image that collects is divided at least two image blocks, from the current image frame that is divided into two image blocks, select the image block that will detect at least according to default image block cycle detection order, and with this image block as area-of-interest.The present invention effectively raises the real-time of classifier algorithm by piecemeal and default image block cycle detection order, has promoted detection speed.
Further, the present invention also comprises the pedestrian zone of further location is followed the tracks of, thereby reduced loss after employing svm classifier device is further located the pedestrian zone, simultaneously, also promoted detection speed indirectly.
Further, apply the present invention to automotive field, utilize pedestrian's image of the camera collection peripheral vehicle that is arranged at body of a motor car, the pedestrian in the image is detected, guaranteed driver's traffic safety, effectively avoided the generation of traffic accident incident.
Further, utilize the pedestrian of the preposition camera collection vehicle front of the automobile image of leaning to one side, the pedestrian that the pedestrian is leaned to one side in the image detects, in the higher position of vehicle front accident rate, can effectively guarantee driver's traffic safety by the present invention, effectively avoid the generation of traffic accident incident.
Description of drawings
Fig. 1 is a kind of pedestrian detection method process flow diagram of the embodiment of the invention;
Fig. 2 is the preposition camera synoptic diagram of embodiment of the invention automobile;
Fig. 3 is an embodiment of the invention image block synoptic diagram;
Fig. 4 is an embodiment of the invention pedestrian regional area synoptic diagram;
Fig. 5 is a kind of pedestrian detection method process flow diagram of another embodiment of the present invention;
Fig. 6 is a kind of pedestrian detection device module map of the embodiment of the invention.
Embodiment
In conjunction with the accompanying drawings the present invention is described in further detail below by embodiment.
A kind of pedestrian detection method comprises:
Gather pedestrian's image, and image is carried out pre-service, from pretreated image, extract area-of-interest; Adopt the pedestrian zone in the Adaboost cascade classifier pre-determined bit area-of-interest; Adopt the svm classifier device in the pedestrian zone of Adaboost cascade classifier pre-determined bit, further to locate the pedestrian zone.
To be applied to automotive field, utilize the pedestrian of the preposition camera collection vehicle front of the automobile image of leaning to one side to be example, Fig. 1 is a kind of pedestrian detection method process flow diagram of the embodiment of the invention, please refer to Fig. 1:
The image of S11, collection vehicle front carries out pre-service to image.Fig. 2 is the preposition camera synoptic diagram of embodiment of the invention automobile, comprise automobile 21 and camera 22, gather the realtime graphic in automobile 21 the place aheads by the preposition camera 22 on the automobile 21 (for example infrared CCD camera or CMOS camera), and image carried out pre-service, for example according to the needs of data layout, the image transitions of obtaining is become the single channel gray level image, image is carried out the histogram equalization operation, reduce the influence of illumination and background; Can also suitably adjust size of images etc.; Estimate based on the positional information that pedestrian in the realtime graphic of gathering may bump, the image that pedestrian and pedestrian be not in the hazardous location can not occur to part and be not counted in detection, for example remove the sky and the ground scene of image top and the bottom pixel, and the road both sides scene of image left and right sides partial pixel, only in more among a small circle, detect the pedestrian, reduce the Flame Image Process area, thereby reduce data processing amount, improve the algorithm real-time.
S12, from pretreated image, extract area-of-interest.The pedestrian's image that collects is divided at least two image blocks, from the current image frame that is divided into two image blocks, selects the image block that will detect at least according to default image block cycle detection order, and with this image block as area-of-interest.
Between present embodiment image block and the image block overlapping region can be arranged, the overlapping region can be the 5-10m size of pedestrian in image far away, selects one of them image block to detect each two field picture according to default image block cycle detection order.Please refer to Fig. 3, Fig. 3 is an embodiment of the invention image block synoptic diagram, the pedestrian's image that collects is divided into n image block, n is more than or equal to 2, to image block n, default image block cycle detection order comprises from image block 1 and detecting to image block n, detects after the image block n such as image block 1, image block 2, next frame is got back to image block 1 cycle detection from image block n again, selects the image block that will detect from current image frame.Equal 3 such as n, the image block of selecting in the former frame image to detect is an image block 1, be in the former frame image with image block 1 as area-of-interest, the image block of selecting in the current frame image to detect is an image block 2, be in the current frame image with image block 2 as area-of-interest, the image block of selecting in the next frame image to detect is an image block 3, be in the next frame image with image block 3 as area-of-interest, the image block of selecting in the next frame image again to detect is again an image block 1, promptly again in the next frame image with image block 1 as area-of-interest, circulation successively.
S13, the area-of-interest that extracts is carried out one times dwindling, narrow down to 1/2 of original size, minimum detection frame size is dwindling in proportion also, under the situation that does not influence testing result, reduced the treatment of picture time, improved the real-time of algorithm, reduction detection block size when follow-up calibration, i.e. expansion is twice.
Pedestrian zone in S14, the employing Adaboost cascade classifier pre-determined bit area-of-interest, the result judges whether that pre-determined bit arrives the pedestrian zone according to the output of Adaboost cascade classifier, if, then carry out S15, simultaneously, the Adaboost cascade classifier judges whether to need to detect the next frame image, if do not need, then finish, if desired, then continue to detect the next frame image; If there is not pre-determined bit to arrive the pedestrian zone, then the Adaboost cascade classifier judges whether to need to detect the next frame image, if do not need, then finishes, and if desired, then continues to detect the next frame image.
The application scenarios of present embodiment is that the pedestrian of vehicle front crosses the street, pedestrian's the attitude of leaning to one side under the special scenes that crosses the street according to the pedestrian, present embodiment is in order to improve detection efficiency, can only detect the pedestrian that leans to one side in the image, the Adaboost cascade classifier of present embodiment can be the Adaboost cascade classifier that adopt a large amount of pedestrians to lean to one side pedestrian that sample and negative sample be trained to leans to one side so, further, because the Adaboost cascade classifier of present embodiment is the pedestrian zone in the pre-determined bit area-of-interest, therefore, in order to improve detection efficiency, the Adaboost cascade classifier of present embodiment is the whole Adaboost cascade classifier of pedestrian preferably, adopt algorithm that lean to one side whole sample and negative sample of a large amount of pedestrians trained, extract Haar feature (rectangular characteristic), the pedestrian who the generates present embodiment whole Adaboost cascade classifier of leaning to one side, according to the pedestrian who has trained the whole Adaboost cascade classifier of leaning to one side, pre-determined bit pedestrian zone in area-of-interest obtains the pedestrian zone of mating most.Concrete grammar is as follows: before the employing feature templates travels through image image is carried out integrogram and calculate, it is described that integrogram is calculated as follows formula:
Wherein (u v) represents the coordinate points in the integrogram, (x, y) represent the coordinate points of original image, (u v) represents (u in the integrogram to S, v) dot product score value, (x y) represents original image (x to I, y) some pixel value, the gray-scale value of every bit equals in the original image to be to the gray scale summation of the rectangular area of angle point, to obtain each eigenwert thus, use the Adaboost algorithm with initial point and this point on the integral image, at feature templates, train 12 grades cascade classifier.Use the pedestrian Haar eigenwert (rectangular characteristic) that writes down in 12 grades the cascade classifier that area-of-interest is detected and locate, thereby obtain the pedestrian zone of coupling, its judgment formula is:
Wherein H (x) is the expression formula of one-level strong classifier, and T represents the Weak Classifier number that the one-level strong classifier is comprised, h
t(x) expression formula of t Weak Classifier of expression, α
tThe weight of representing t Weak Classifier, calculated by the image of 12 grades cascade classifiers to area-of-interest, if each level is all exported H (x)=1, then representative navigates to the pedestrian zone, if wherein one-level output H (x)=0 is arranged, then representative does not detect the pedestrian zone.
The Adaboost cascade classifier of present embodiment is not limited to pedestrian that adopt a large amount of pedestrians to lean to one side the pedestrian that leans to one side in the image that sample and negative sample be trained to the detects Adaboost cascade classifier of leaning to one side, if want to detect the identity pedestrian in the image, the Adaboost cascade classifier that the identity pedestrian in the image that can adopt a large amount of pedestrian's identity samples and negative sample to be trained to detects; Simultaneously, the whole Adaboost cascade classifier of pedestrian that the pedestrian's integral body in the image that adopts whole sample of a large amount of pedestrians and negative sample to be trained to of being not limited to the Adaboost cascade classifier of present embodiment detects, Adaboost cascade classifier in the present embodiment can also be the local Adaboost cascade classifier of pedestrian that the pedestrian part in the image of adopting a large amount of pedestrian's fractional sample and negative sample to be trained to is detected, or adopt a large amount of pedestrians to lean to one side the pedestrian in the image that fractional sample and negative sample be trained to pedestrian that the part the detects local Adaboost cascade classifier of leaning to one side of leaning to one side.
S15, employing svm classifier device are further located the pedestrian zone in the pedestrian zone of Adaboost cascade classifier pre-determined bit, the result judges whether further to navigate to the pedestrian zone according to the output of svm classifier device, if, then the pedestrian zone that obtains is shown to the user, simultaneously, the svm classifier device judges whether to need to detect the next frame image, if do not need, then finish, if desired, then continue to detect the next frame image; If further the location does not obtain the pedestrian zone, then the svm classifier device judges whether to need to detect the next frame image, if do not need, then finishes, and if desired, then continues to detect the next frame image.
Because there was more flase drop in the pedestrian zone of pre-determined bit during previous step was rapid, by feature extraction is carried out in these zones, the feature of extracting is judged according to the svm classifier device that has trained, remove the flase drop zone, thus pedestrian zone, more accurate location.Because the svm classifier device of present embodiment is that the pedestrian zone of pre-determined bit in the previous step is further located, therefore, in order to improve the accuracy of detection, the svm classifier device of present embodiment comprises whole svm classifier device of pedestrian and the local svm classifier device of pedestrian, can be according to actual conditions, select one or both in conjunction with coming pedestrian zone further to locate to pre-determined bit, simultaneously, in order to cooperate the application scenarios of present embodiment, the pedestrian who is vehicle front crosses the street, can only detect the pedestrian that leans to one side in the image, improve detection efficiency, the svm classifier device of present embodiment can be adopt a large amount of pedestrians to lean to one side pedestrian that whole sample and negative sample the be trained to whole svm classifier device of leaning to one side so, and adopt a large amount of pedestrians to lean to one side pedestrian that fractional sample and negative sample the be trained to local svm classifier device of leaning to one side.The lean to one side training method of whole svm classifier device of the pedestrian of present embodiment is as follows: lean to one side whole sample and negative sample of a large amount of pedestrians carried out HOG feature (gradient orientation histogram feature) and extract, use training tool to obtain SVM weights and threshold value, generate pedestrian's whole svm classifier device of leaning to one side, the HOG feature is to describe the proper vector of given image window gradient statistical information, in order better feature extraction to be carried out in zone to be checked, all zones to be checked can be enlarged or are contracted to identical size, 64*128 size for example, calculate its HOG proper vector, the piece that comprises 105 16*16 sizes such as the detection block of 64*128 size, piece and interblock have overlapping, each piece is expressed as the feature of 36 dimensions, whole detection block is described with 3780 proper vector, utilize the SVM algorithm, at the HOG feature, training side pedestrian's integral body and local support vector machine svm classifier device, sorter provides weight w and threshold value b, the pedestrian zone of pre-determined bit is judged with the 3780 dimension HOG feature x that asked its judgment formula is according to the support vector machine svm classifier device that has trained:
Wherein w and b are respectively the weight and the threshold value of sorter, and x represents the HOG proper vector in the pre-determined bit zone imported, and n is the support vector number of sorter B, (x
i, y
i) i support vector of expression, λ
iBe Lagrangian coefficient, if f (x)=1, then expression detects and is the pedestrian.The lean to one side training method of local svm classifier device of the pedestrian of present embodiment is as follows: lean to one side fractional sample and negative sample of a large amount of pedestrians carried out HOG feature (gradient orientation histogram feature) and extract, use training tool to obtain SVM weights and threshold value, generate pedestrian's local svm classifier device of leaning to one side, pedestrian's regional area comprises the pedestrian head zone, pedestrian's shoulder regions, pedestrian's trunk zone, in the pedestrian lower leg zone one or more, corresponding pedestrian's fractional sample comprises the pedestrian head sample, pedestrian's shoulder sample, pedestrian's trunk sample, in the pedestrian lower leg sample one or more.Fig. 4 is an embodiment of the invention pedestrian regional area synoptic diagram, please refer to Fig. 4: present embodiment is according to concrete application scenarios, the pedestrian is divided into three zones, it is respectively a shoulder 41, trunk 42 and leg area 43, have overlapping between the adjacent area, respectively to these three zones in conjunction with specifying pedestrians' fractional sample of leaning to one side to train, obtain this trizonal pedestrian local svm classifier device of leaning to one side respectively, shoulder is made screening fast by " recessed " font feature to the pedestrian zone of pre-determined bit or is judged, trunk is made screening fast by " wood " font feature to the pedestrian zone of pre-determined bit or is judged, shank is made screening fast by " people " font feature to the pedestrian zone of pre-determined bit or is judged, with the leg area is example, use the window size of 50*50, the block size of 20*20, the step-length of piece and interblock is 10, always have 16 pieces, each piece is expressed as the proper vector of 36 dimensions, so the total characteristic dimension of window is the 16*36=576 dimension, comparing whole 3780 dimensional feature vectors of pedestrian significantly reduces, thereby, improved real-time to the characteristics that the pedestrian who crosses the street has efficient screening.By shank sample and negative sample calculated characteristics vector that a large amount of pedestrians are leaned to one side, utilize SVM training to draw pedestrian's shank svm classifier device of leaning to one side, window with 50*50 travels through all frames to be detected, use this svm classifier device to judge, if function is output as 1, then representative detects people's leg, if be output as 0, does not then detect.
The svm classifier device of present embodiment is not limited to pedestrian's whole svm classifier device of leaning to one side, and pedestrian's local svm classifier device of leaning to one side, if want to detect the identity pedestrian in the image, the svm classifier device that the identity pedestrian in the image that can adopt a large amount of pedestrian's identity samples and negative sample to be trained to detects;
The method of extracting area-of-interest from pretreated image has multiple, be not limited to a kind of mode cited among the present embodiment S12, such as adopting mode habitual in the prior art, be about to pretreated integral image as area-of-interest, and it does not carried out piecemeal.
Further, present embodiment comprises that also Fig. 5 is a kind of pedestrian detection method process flow diagram of another embodiment of the present invention, please refer to Fig. 5 to the further pedestrian zone that obtains, the location step of following the tracks of among the S15:
Step S11 to S15 is as above-mentioned, and the svm classifier device further obtains after the pedestrian zone location in S15, when the pedestrian zone that obtains is shown to the user, carries out S16.
Follow the tracks of in S16, the pedestrian zone that further location is obtained, and carries out following steps: will further locate the pedestrian zone that obtains and enlarge according to setting size or ratio;
When from the next frame image, extracting area-of-interest, with the pedestrian zone that enlarges in the next frame image institute corresponding region and the image block that will detect from the next frame picture frame, selected according to image block cycle detection order together as area-of-interest.
Such as, the image block of selecting in the current frame image to detect is an image block 2, be in the current frame image with image block 2 as area-of-interest, pedestrian zone after step S13 to S16 is further located, enlarge according to setting size or ratio further locating the pedestrian zone that obtains in the current frame image, the pedestrian zone that enlarges can greater than, be equal to or less than the size of image block 2, also can be with image block 2 whole amplifications, when from the next frame image, extracting area-of-interest, with the pedestrian zone that enlarges in the next frame image institute corresponding region and the image block (being image block 3) of from the next frame image, selecting detection according to image block cycle detection order together as the area-of-interest of next frame image.
Adopt support vector machine svm classifier device in the pedestrian zone of Adaboost cascade classifier pre-determined bit, further to locate after the pedestrian zone, comprise that also the pedestrian zone that further location is obtained follows the tracks of, enlarge according to setting size or ratio further locating the pedestrian zone that obtains, when from the next frame image, extracting area-of-interest, with the institute corresponding region in the next frame image, pedestrian zone that enlarges, with the image block of from the next frame picture frame, selecting according to image block cycle detection order that will detect together as area-of-interest, can prevent omission like this, promote the accuracy that detects, simultaneously the real-time of boosting algorithm.
Fig. 6 is the module map of a kind of pedestrian detection device of the embodiment of the invention, please refer to Fig. 6: a kind of pedestrian detection device, comprise image capture module 61, image processing module 62, Adaboost cascade classifier 63 and svm classifier device 64, wherein, image capture module 61 is used to gather pedestrian's image, image processing module 62 is used for the image that collects is carried out pre-service, from pretreated image, extract area-of-interest, Adaboost cascade classifier 63 is used for the pedestrian zone of pre-determined bit area-of-interest, and svm classifier device 64 is used for further locating the pedestrian zone in the pedestrian zone of Adaboost cascade classifier 63 pre-determined bit.Adaboost cascade classifier 63 comprises Adaboost cascade classifier that employing pedestrian global feature is trained to and/or the Adaboost cascade classifier that adopts pedestrian's local feature white silk to become; Svm classifier device 64 comprises svm classifier device that employing pedestrian global feature is trained to and/or the svm classifier device that adopts pedestrian's local feature white silk to become.
Further, pedestrian's image that image processing module 62 specifically is used for collecting is divided at least two image blocks, from the current image frame that is divided into two image blocks, select the image block that will detect at least according to default image block cycle detection order, and with this image block as area-of-interest.
Further, image processing module 62 also is used for adopting the svm classifier device after the pedestrian zone is further located in the pedestrian zone of Adaboost cascade classifier pre-determined bit, the pedestrian zone that further location obtains is followed the tracks of, comprise: judge whether to navigate to the pedestrian zone, if not, continue to gather pedestrian's image and from described image, extract area-of-interest, if, then carry out following steps: will further locate the pedestrian zone that obtains and enlarge according to setting size or ratio, when from the next frame image, extracting area-of-interest, with the pedestrian zone that enlarges in the next frame image institute corresponding region and the image block that will detect from the next frame picture frame, selected according to image block cycle detection order together as area-of-interest.
Further, image capture module 62 comprises the camera that is arranged at body of a motor car, is used to gather pedestrian's image of peripheral vehicle, especially is arranged at the camera of vehicle front, the pedestrian who is used to the to gather vehicle front image of leaning to one side.
The Adaboost cascade classifier that present embodiment carries out pre-determined bit to the pedestrian zone is pedestrian's whole tagsort device of leaning to one side under the special scenes that crosses the street of the suitable pedestrian of training in advance, to pedestrian zone further the svm classifier device of location be lean to one side integral body and local tagsort device of the pedestrian of training in advance, the Adaboos t cascade classifier of present embodiment have detection speed soon, stronger characteristics such as universality, the svm classifier utensil has the accuracy of height, and the local svm classifier device of employing pedestrian accuracy is higher, speed is faster.Present embodiment is applied to automotive field, blured the personal feature between the vehicle front pedestrian, reduced the influence to testing result such as individuality difference, illumination, reduced false drop rate, improved detection efficiency, further, by the pedestrian's attitude of crossing road on the sorter detection road that adopts side pedestrian's local feature, at specific application scenarios, reduced the non-influence of crossing the road pedestrian of different attitudes to testing result, improved the validity and the real-time of algorithm.Because the present vehicles and road equipment are more and more flourishing, the complexity of traffic also increases thereupon greatly, and driver's information judgement is more and more limited to comparatively speaking, so just, caused the increasing of raising of traffic hazard incidence and hidden danger risk, present embodiment utilizes pedestrian's image of the preposition camera collection vehicle front of automobile, pedestrian in the image is detected, guaranteed driver's traffic safety, effectively avoided the generation of traffic accident incident.
Above content be in conjunction with concrete embodiment to further describing that the present invention did, can not assert that concrete enforcement of the present invention is confined to these explanations.For the general technical staff of the technical field of the invention, without departing from the inventive concept of the premise, can also make some simple deduction or replace, all should be considered as belonging to protection scope of the present invention.
Claims (10)
1. a pedestrian detection method is characterized in that, comprising:
Gather pedestrian's image, from described image, extract area-of-interest;
Adopt the pedestrian zone in the described area-of-interest of Adaboost cascade classifier pre-determined bit;
Adopt support vector machine svm classifier device in the pedestrian zone of described Adaboost cascade classifier pre-determined bit, further to locate the pedestrian zone.
2. the method for claim 1, it is characterized in that described Adaboost cascade classifier comprises whole Adaboost cascade classifier of pedestrian that the whole sample of employing pedestrian becomes with the respective negative sample training and/or the local Adaboost cascade classifier of pedestrian that adopts pedestrian's fractional sample to become with the respective negative sample training; Described svm classifier device comprises whole svm classifier device of pedestrian that the whole sample of employing pedestrian becomes with the respective negative sample training and/or the local svm classifier device of pedestrian that adopts pedestrian's fractional sample to become with the respective negative sample training.
3. method as claimed in claim 1 or 2, it is characterized in that, gather pedestrian's image, the method of extracting area-of-interest from described image is specially: the pedestrian's image that collects is divided at least two image blocks, from the current image frame that is divided into two image blocks, select the image block that will detect at least according to default image block cycle detection order, and with this image block as area-of-interest.
4. method as claimed in claim 3, it is characterized in that, adopt support vector machine svm classifier device in the pedestrian zone of described Adaboost cascade classifier pre-determined bit further after the pedestrian zone, location, also comprise the step that follow the tracks of in pedestrian zone that further location is obtained, comprising:
Judge whether to navigate to the pedestrian zone, enlarge according to setting size or ratio if then will further locate the pedestrian zone that obtains;
When from the next frame image, extracting area-of-interest, with the pedestrian zone that enlarges in the next frame image institute corresponding region and the image block that will detect from the next frame picture frame, selected according to image block cycle detection order together as area-of-interest.
5. method as claimed in claim 1 or 2 is characterized in that, the mode of gathering pedestrian's image comprises: the pedestrian's image that utilizes the camera collection peripheral vehicle that is arranged at body of a motor car.
6. method as claimed in claim 5 is characterized in that, utilizes the pedestrian of the preposition camera collection vehicle front of the automobile image of leaning to one side.
7. a pedestrian detection device is characterized in that, comprises image capture module, image processing module, Adaboost cascade classifier and support vector machine svm classifier device, wherein,
Described image capture module is used to gather pedestrian's image;
Described image processing module is used for extracting area-of-interest from described image;
Described Adaboost cascade classifier is used for the pedestrian zone of the described area-of-interest of pre-determined bit;
Described support vector machine svm classifier device is used for further locating the pedestrian zone in the pedestrian zone of described Adaboost cascade classifier pre-determined bit.
8. device as claimed in claim 7, it is characterized in that described Adaboost cascade classifier comprises whole Adaboost cascade classifier of pedestrian that the whole sample of employing pedestrian becomes with the respective negative sample training and/or the local Adaboost cascade classifier of pedestrian that adopts pedestrian's fractional sample to become with respective negative sample white silk; Described svm classifier device comprises whole svm classifier device of pedestrian that the whole sample of employing pedestrian becomes with the respective negative sample training and/or the local svm classifier device of pedestrian that adopts pedestrian's fractional sample to become with respective negative sample white silk.
9. as claim 7 or 8 described devices, it is characterized in that, pedestrian's image that described image processing module specifically is used for collecting is divided at least two image blocks, from the current image frame that is divided into two image blocks, select the image block that will detect at least according to default image block cycle detection order, and with this image block as area-of-interest.
10. as claim 7 or 8 described devices, it is characterized in that described image capture module comprises the camera that is arranged at body of a motor car, be used to gather pedestrian's image of peripheral vehicle.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201110132332 CN102201059A (en) | 2011-05-20 | 2011-05-20 | Pedestrian detection method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201110132332 CN102201059A (en) | 2011-05-20 | 2011-05-20 | Pedestrian detection method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN102201059A true CN102201059A (en) | 2011-09-28 |
Family
ID=44661722
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN 201110132332 Pending CN102201059A (en) | 2011-05-20 | 2011-05-20 | Pedestrian detection method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102201059A (en) |
Cited By (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102496001A (en) * | 2011-11-15 | 2012-06-13 | 无锡港湾网络科技有限公司 | Method of video monitor object automatic detection and system thereof |
| CN102609682A (en) * | 2012-01-13 | 2012-07-25 | 北京邮电大学 | Feedback pedestrian detection method for region of interest |
| CN102609716A (en) * | 2012-01-10 | 2012-07-25 | 银江股份有限公司 | Pedestrian detecting method based on improved HOG feature and PCA (Principal Component Analysis) |
| CN103177263A (en) * | 2013-03-13 | 2013-06-26 | 浙江理工大学 | Image-based automatic detection and counting method for rice field planthopper |
| CN103473953A (en) * | 2013-08-28 | 2013-12-25 | 奇瑞汽车股份有限公司 | Pedestrian detection method and system |
| CN103530646A (en) * | 2012-06-12 | 2014-01-22 | 通用汽车环球科技运作有限责任公司 | Complex-object detection using a cascade of classifiers |
| CN103605966A (en) * | 2013-11-26 | 2014-02-26 | 奇瑞汽车股份有限公司 | Method and device for identifying pedestrians |
| CN103646257A (en) * | 2013-12-30 | 2014-03-19 | 中国科学院自动化研究所 | Video monitoring image-based pedestrian detecting and counting method |
| CN103661102A (en) * | 2012-08-31 | 2014-03-26 | 北京旅行者科技有限公司 | Method and device for reminding passersby around vehicles in real time |
| CN104952060A (en) * | 2015-03-19 | 2015-09-30 | 杭州电子科技大学 | Adaptive segmentation extraction method for infrared pedestrian region of interests |
| CN105023001A (en) * | 2015-07-17 | 2015-11-04 | 武汉大学 | Selective region-based multi-pedestrian detection method and system |
| CN105095869A (en) * | 2015-07-24 | 2015-11-25 | 深圳市佳信捷技术股份有限公司 | Pedestrian detection method and apparatus |
| CN105426928A (en) * | 2014-09-19 | 2016-03-23 | 无锡慧眼电子科技有限公司 | Pedestrian detection method based on Haar characteristic and EOH characteristic |
| CN105718904A (en) * | 2016-01-25 | 2016-06-29 | 大连楼兰科技股份有限公司 | Blind people detection and identification method and system based on combined characteristics and vehicle-mounted cameras |
| CN105959639A (en) * | 2016-06-06 | 2016-09-21 | 南京工程学院 | Pedestrian monitoring method in urban street area based on ground calibration |
| CN106203361A (en) * | 2016-07-15 | 2016-12-07 | 苏州宾果智能科技有限公司 | A kind of robotic tracking's method and apparatus |
| CN106326853A (en) * | 2016-08-19 | 2017-01-11 | 厦门美图之家科技有限公司 | Human face tracking method and device |
| CN106845352A (en) * | 2016-12-23 | 2017-06-13 | 北京旷视科技有限公司 | Pedestrian detection method and device |
| CN106980814A (en) * | 2016-01-15 | 2017-07-25 | 福特全球技术公司 | With the pedestrian detection of conspicuousness map |
| CN107358149A (en) * | 2017-05-27 | 2017-11-17 | 深圳市深网视界科技有限公司 | A kind of human body attitude detection method and device |
| CN104512327B (en) * | 2013-09-27 | 2017-11-21 | 比亚迪股份有限公司 | Blind area vehicle checking method, system, vehicle lane change method for early warning and system |
| CN108154076A (en) * | 2017-11-16 | 2018-06-12 | 北京遥感设备研究所 | A kind of cascade rail pedestrian detection method of machine learning algorithm |
| CN109063651A (en) * | 2018-08-06 | 2018-12-21 | 百度在线网络技术（北京）有限公司 | object detecting method, device, equipment and storage medium |
| CN109410582A (en) * | 2018-11-27 | 2019-03-01 | 易念科技（深圳）有限公司 | Traffic condition analysis method and terminal device |
| CN109919067A (en) * | 2019-02-27 | 2019-06-21 | 智慧海派科技有限公司 | A kind of image recognition method and safe automatic Pilot method |
| CN112183206A (en) * | 2020-08-27 | 2021-01-05 | 广州中国科学院软件应用技术研究所 | Traffic participant positioning method and system based on roadside monocular camera |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101075265A (en) * | 2007-07-23 | 2007-11-21 | 王文钢 | Method for demonstrating webpage window video in automatically-tracting picture emphasis zone |
| CN101369315A (en) * | 2007-08-17 | 2009-02-18 | 上海银晨智能识别科技有限公司 | Human face detection method |
| CN101714213A (en) * | 2009-11-27 | 2010-05-26 | 中国科学技术大学 | Vehicle-mounted pedestrian detection method and system |
| CN101887524A (en) * | 2010-07-06 | 2010-11-17 | 湖南创合制造有限公司 | Pedestrian detection method based on video monitoring |
-
2011
- 2011-05-20 CN CN 201110132332 patent/CN102201059A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101075265A (en) * | 2007-07-23 | 2007-11-21 | 王文钢 | Method for demonstrating webpage window video in automatically-tracting picture emphasis zone |
| CN101369315A (en) * | 2007-08-17 | 2009-02-18 | 上海银晨智能识别科技有限公司 | Human face detection method |
| CN101714213A (en) * | 2009-11-27 | 2010-05-26 | 中国科学技术大学 | Vehicle-mounted pedestrian detection method and system |
| CN101887524A (en) * | 2010-07-06 | 2010-11-17 | 湖南创合制造有限公司 | Pedestrian detection method based on video monitoring |
Cited By (38)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102496001B (en) * | 2011-11-15 | 2015-02-25 | 无锡港湾网络科技有限公司 | Method of video monitor object automatic detection and system thereof |
| CN102496001A (en) * | 2011-11-15 | 2012-06-13 | 无锡港湾网络科技有限公司 | Method of video monitor object automatic detection and system thereof |
| CN102609716A (en) * | 2012-01-10 | 2012-07-25 | 银江股份有限公司 | Pedestrian detecting method based on improved HOG feature and PCA (Principal Component Analysis) |
| CN102609682B (en) * | 2012-01-13 | 2014-02-05 | 北京邮电大学 | Feedback pedestrian detection method for region of interest |
| CN102609682A (en) * | 2012-01-13 | 2012-07-25 | 北京邮电大学 | Feedback pedestrian detection method for region of interest |
| CN103530646A (en) * | 2012-06-12 | 2014-01-22 | 通用汽车环球科技运作有限责任公司 | Complex-object detection using a cascade of classifiers |
| CN103530646B (en) * | 2012-06-12 | 2017-08-15 | 通用汽车环球科技运作有限责任公司 | The complex object cascaded using grader is detected |
| CN103661102A (en) * | 2012-08-31 | 2014-03-26 | 北京旅行者科技有限公司 | Method and device for reminding passersby around vehicles in real time |
| CN103177263A (en) * | 2013-03-13 | 2013-06-26 | 浙江理工大学 | Image-based automatic detection and counting method for rice field planthopper |
| CN103177263B (en) * | 2013-03-13 | 2016-03-23 | 浙江理工大学 | A kind of rice field plant hopper based on image detects and method of counting automatically |
| 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 |
| CN103473953A (en) * | 2013-08-28 | 2013-12-25 | 奇瑞汽车股份有限公司 | Pedestrian detection method and system |
| CN104512327B (en) * | 2013-09-27 | 2017-11-21 | 比亚迪股份有限公司 | Blind area vehicle checking method, system, vehicle lane change method for early warning and system |
| CN103605966B (en) * | 2013-11-26 | 2017-01-11 | 奇瑞汽车股份有限公司 | Method and device for identifying pedestrians |
| CN103605966A (en) * | 2013-11-26 | 2014-02-26 | 奇瑞汽车股份有限公司 | Method and device for identifying pedestrians |
| CN103646257B (en) * | 2013-12-30 | 2017-06-16 | 中国科学院自动化研究所 | A kind of pedestrian detection and method of counting based on video monitoring image |
| CN103646257A (en) * | 2013-12-30 | 2014-03-19 | 中国科学院自动化研究所 | Video monitoring image-based pedestrian detecting and counting method |
| CN105426928A (en) * | 2014-09-19 | 2016-03-23 | 无锡慧眼电子科技有限公司 | Pedestrian detection method based on Haar characteristic and EOH characteristic |
| CN105426928B (en) * | 2014-09-19 | 2019-06-18 | 江苏慧眼数据科技股份有限公司 | A kind of pedestrian detection method based on Haar feature and EOH feature |
| CN104952060B (en) * | 2015-03-19 | 2017-10-31 | 杭州电子科技大学 | A kind of infrared pedestrian's area-of-interest adaptivenon-uniform sampling extracting method |
| CN104952060A (en) * | 2015-03-19 | 2015-09-30 | 杭州电子科技大学 | Adaptive segmentation extraction method for infrared pedestrian region of interests |
| CN105023001B (en) * | 2015-07-17 | 2018-03-27 | 武汉大学 | A kind of more pedestrian detection methods and system based on selective area |
| CN105023001A (en) * | 2015-07-17 | 2015-11-04 | 武汉大学 | Selective region-based multi-pedestrian detection method and system |
| CN105095869A (en) * | 2015-07-24 | 2015-11-25 | 深圳市佳信捷技术股份有限公司 | Pedestrian detection method and apparatus |
| CN106980814A (en) * | 2016-01-15 | 2017-07-25 | 福特全球技术公司 | With the pedestrian detection of conspicuousness map |
| CN105718904A (en) * | 2016-01-25 | 2016-06-29 | 大连楼兰科技股份有限公司 | Blind people detection and identification method and system based on combined characteristics and vehicle-mounted cameras |
| CN105959639A (en) * | 2016-06-06 | 2016-09-21 | 南京工程学院 | Pedestrian monitoring method in urban street area based on ground calibration |
| CN105959639B (en) * | 2016-06-06 | 2019-06-14 | 南京工程学院 | Pedestrian's monitoring method in avenue region based on ground calibration |
| CN106203361A (en) * | 2016-07-15 | 2016-12-07 | 苏州宾果智能科技有限公司 | A kind of robotic tracking's method and apparatus |
| CN106326853A (en) * | 2016-08-19 | 2017-01-11 | 厦门美图之家科技有限公司 | Human face tracking method and device |
| CN106845352A (en) * | 2016-12-23 | 2017-06-13 | 北京旷视科技有限公司 | Pedestrian detection method and device |
| CN107358149A (en) * | 2017-05-27 | 2017-11-17 | 深圳市深网视界科技有限公司 | A kind of human body attitude detection method and device |
| CN108154076A (en) * | 2017-11-16 | 2018-06-12 | 北京遥感设备研究所 | A kind of cascade rail pedestrian detection method of machine learning algorithm |
| CN109063651A (en) * | 2018-08-06 | 2018-12-21 | 百度在线网络技术（北京）有限公司 | object detecting method, device, equipment and storage medium |
| CN109410582A (en) * | 2018-11-27 | 2019-03-01 | 易念科技（深圳）有限公司 | Traffic condition analysis method and terminal device |
| CN109919067A (en) * | 2019-02-27 | 2019-06-21 | 智慧海派科技有限公司 | A kind of image recognition method and safe automatic Pilot method |
| CN112183206A (en) * | 2020-08-27 | 2021-01-05 | 广州中国科学院软件应用技术研究所 | Traffic participant positioning method and system based on roadside monocular camera |
| CN112183206B (en) * | 2020-08-27 | 2024-04-05 | 广州中国科学院软件应用技术研究所 | Traffic participant positioning method and system based on road side monocular camera |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102201059A (en) | Pedestrian detection method and device | |
| US9842266B2 (en) | Method for detecting driver cell phone usage from side-view images | |
| WO2019196130A1 (en) | Classifier training method and device for vehicle-mounted thermal imaging pedestrian detection | |
| Gerónimo et al. | 2D–3D-based on-board pedestrian detection system | |
| CN101477626B (en) | Method for detecting human head and shoulder in video of complicated scene | |
| CN101211408B (en) | Vehicle side image recognition method and apparatus, car lamp error identification detection and driving safety prediction method | |
| US8131079B2 (en) | Pedestrian detection device and pedestrian detection method | |
| US9626599B2 (en) | Reconfigurable clear path detection system | |
| CN104091171B (en) | Vehicle-mounted far infrared pedestrian detecting system and method based on local feature | |
| US20150286883A1 (en) | Robust windshield detection via landmark localization | |
| CN102385703B (en) | A kind of identity identifying method based on face and system | |
| CN102768726B (en) | Pedestrian detection method for preventing pedestrian collision | |
| CN101980245B (en) | Adaptive template matching-based passenger flow statistical method | |
| Alonso et al. | Robust vehicle detection through multidimensional classification for on board video based systems | |
| CN102682304A (en) | Multi-feature integrated passer-by detection method and device | |
| RU2012122205A (en) | METHOD AND DEVICE FOR DETECTING DRIVING UNDER TIREDNESS AND A CAR USING THESE METHOD AND DEVICE | |
| CN102867188A (en) | Method for detecting seat state in meeting place based on cascade structure | |
| CN103632170A (en) | Pedestrian detection method and device based on characteristic combination | |
| CN104268596A (en) | License plate recognizer and license plate detection method and system thereof | |
| CN104268589B (en) | A kind of front vehicles detection method | |
| EP2741234B1 (en) | Object localization using vertical symmetry | |
| CN104463232A (en) | Density crowd counting method based on HOG characteristic and color histogram characteristic | |
| CN103679205A (en) | Preceding car detection method based on shadow hypothesis and layered HOG (histogram of oriented gradient) symmetric characteristic verification | |
| CN102609724A (en) | Method for prompting ambient environment information by using two cameras | |
| CN101470802A (en) | Object detection apparatus and method thereof |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C12 | Rejection of a patent application after its publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20110928 |