CN102768726B - Pedestrian detection method for preventing pedestrian collision - Google Patents
Pedestrian detection method for preventing pedestrian collision Download PDFInfo
- Publication number
- CN102768726B CN102768726B CN201110116569.0A CN201110116569A CN102768726B CN 102768726 B CN102768726 B CN 102768726B CN 201110116569 A CN201110116569 A CN 201110116569A CN 102768726 B CN102768726 B CN 102768726B
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- image
- region
- sorter
- feature
- 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.)
- Active
Links
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a pedestrian detection method for preventing pedestrian collision. The method comprises the following steps: collecting images in front of a vehicle and pre-processing the images; extracting interesting zones of pedestrians; blocking the images and re-setting the interesting zones; zooming out the to-be-detected zones of each frame of image by one time; pre-localizing the pedestrians in the blocked interesting zones; judging the pre-localized pedestrian zones; and tracking the detected pedestrians. The detection method provided by the invention detects the pedestrians on roads by adopting a pedestrian sorter to fuzz up the individual characteristics of the pedestrians, reduce the effects of the individual differences on the detection result, also reduce the effects of light on the detection result, and enhance the pedestrian detection efficiency; in the meantime, with the adoption of the side pedestrian posture sorter, the pedestrians crossing a road are detected; aiming at the special application scene, the effects of the pedestrians with different postures who do not cross the road on the judgment is reduced; and the effectiveness and the real time property of the algorithm for judging the pedestrians are enhanced.
Description
Technical field
The present invention, about a kind of pedestrian detection method, refers to a kind of pedestrian detection method of the prevention pedestrian impact based on machine vision especially.
Background technology
Image processing technique is utilized to become the megatrend of Automobile Design for automobile supplementary security system increases function.Under the regulation of European Union drives, the headstock profile of automobile and structural design have considered when with pedestrian collision, reduce injury and the mortality ratio of pedestrian as far as possible.At present, utilize profile and structural design to reduce this method of casualties and obtained certain effect, then it is smaller to continue to use this method institute's resulting improvement space.Otherwise, utilize image processing technique to detect pedestrian and prompting sent to driver and even get involved control loop and avoid pedestrian and compare the method for keeping forging ahead, and have very large space to improve existing pedestrian protection system.
Existing pedestrian detecting system generally comprises two modules: region of interest regional partition and target identification.The object of region of interest regional partition from image, extracts the window area that may comprise pedestrian do checking further, to avoid exhaustive search, improves the speed of system.Target identification is the core of pedestrian detecting system, and it is verified the area-of-interest obtained, and to judge wherein whether comprise pedestrian, its performance determines the accuracy of detection and robustness that whole system can reach.At present, pedestrian detection technology generally has following several mode: one, based drive method; Two, based on the method for clear and definite manikin; Three, based on the method for template matches; Four, the method for Corpus--based Method classification.The ultimate principle of above several method and strengths and weaknesses analysis are distinguished as follows:
Based drive method, its principle is that the periodicity by analyzing pedestrian's gait identifies pedestrian; Advantage is less by the impact of color, illumination, compares robust; Shortcoming to identify motion pedestrian, needs multiframe, affect real-time.
Based on the method for clear and definite manikin, its principle constructs clear and definite human parameters model to represent pedestrian; Advantage has clear and definite model, convenient process attitude and occlusion issue; Shortcoming is modeling and solves more complicated.
Based on the method for template matches, its principle is by template representation pedestrian; Advantage is that computing method are simple, and system overhead is little; Shortcoming needs a lot of template to tackle pose problem, and matching ratio is time-consuming comparatively.
The method of Corpus--based Method classification, its principle is identified pedestrian by sorter; Advantage be not needing, quantity of parameters is manually set, robustness is good; Shortcoming be need a large amount of training datas and cycle of training longer.
Summary of the invention
In view of this, fundamental purpose of the present invention is to provide a kind of for automobile, based on the pedestrian detection method of the prevention pedestrian impact of image processing.
For achieving the above object, the invention provides a kind of pedestrian detection method preventing pedestrian impact, it includes following steps:
(1) gather vehicle front image and carry out Image semantic classification;
(2) area-of-interest of pedestrian is extracted;
(3) piecemeal is carried out to image, reset area-of-interest;
(4) one times reducing is carried out to the region to be detected of each two field picture;
(5) in the area-of-interest after piecemeal, pre-determined bit is carried out to pedestrian;
(6) the pedestrian region of pre-determined bit is judged, accurately to locate pedestrian region;
(7) pedestrian detected is followed the tracks of.
In described step (1), by being arranged on the camera collection front realtime graphic on automobile, described Image semantic classification comprises and removes serrating operation and histogram equalization operation to image.
Described step (2) estimates based on pedestrian's positional information that may collide in the realtime graphic gathered, and can not occur that the image that pedestrian and pedestrian are not in hazardous location is not counted in detection, to set the area-of-interest of pedestrian to part.
Described step (3) is that image is divided into equal some pieces, has overlapping region between block and block, only processes wherein one piece of region to each frame input picture, each block region of circular treatment.
Fisrt feature sorter pre-determined bit pedestrian region in pedestrian's area-of-interest is utilized in described step (5), the pedestrian region of second feature sorter to pre-determined bit is utilized to screen and judge in described step (6), accurately to locate pedestrian region, described fisrt feature sorter is the feature classifiers of the attitude information of pedestrian side under the special scenes that crosses the street of the applicable pedestrian of training in advance, and described second feature sorter is the feature classifiers of the overall and local of the pedestrian of training in advance.
Described fisrt feature sorter is rectangular characteristic sorter, and described second feature sorter is HOG feature classifiers.
Adopt AdaBoost algorithm in described step (5), in described step (6), adopt SVM algorithm.
In described step (5), integrogram calculating is carried out to image.
The present invention is by adopting the pedestrian on pedestrian detection of classifier road, and the personal feature between fuzzy pedestrian, decreases the impact of individuality difference on testing result, decrease the impact of illumination on testing result simultaneously, improve the detection efficiency of pedestrian.Simultaneously by adopting the pedestrian of crossing road on the pedestrian's attitude detection of classifier road of side, for specific application scenarios, decreasing the non-impact of crossing road pedestrian and differentiating of different attitude, improve algorithm and differentiate row human effectiveness and real-time.
Accompanying drawing explanation
Fig. 1 is the process flow diagram that the present invention prevents the pedestrian detection method of pedestrian impact.
Embodiment
Below by embodiment, the present invention is described in further detail.
The present invention obtains the realtime graphic of vehicle front by the camera be installed on above automobile, require to extract area-of-interest in the picture according to prevention pedestrian impact, then a series of image procossing and computing are carried out to extracted region, realize the detection of pedestrian and pre-anticollision.
Enforcement of the present invention comprises the following steps:
Step one, gather vehicle front image, front realtime graphic is gathered by the camera (such as infrared CCD camera or CMOS camera) be arranged on automobile, and image is suitably processed, such as according to the needs of data layout, convert the image of acquisition to single channel gray level image, to meet the requirement of database format, serrating operation and histogram equalization operation are gone to image.
Step 2, extract the area-of-interest of pedestrian, the positional information that may collide based on pedestrian in the realtime graphic gathered is estimated, can not occur that the image that pedestrian and pedestrian are not in hazardous location is not counted in detection to part, such as remove sky and the ground scene of image top and the bottom pixel, and the road both sides scene of image left-right parts pixel, thus the area-of-interest of setting pedestrian, only in more among a small circle, search for pedestrian, reduce image procossing area, thus minimizing data processing amount, improve algorithm real-time.
Step 3, carries out piecemeal to image, resets area-of-interest.Image is divided into equal some pieces, has overlapping region between block and block, overlapping region is roughly 5-10m pedestrian far away size in the picture.Wherein one piece of region is only processed to each frame input picture, each block of circular treatment.The method substantially increases the real-time of algorithm.As shown in Figure 1, pretreated image is divided into two pieces of A and B, overlapping C can be had between block with block, two pieces of regions are respectively block A+C and block B+C, each frame detect only to wherein one piece carry out, each block of cycle detection, follows the tracks of the pedestrian detected in the area-of-interest after piecemeal, effectively raises the real-time of algorithm.
Step 4, carries out one times reducing to the region to be detected of each two field picture, i.e. area reduction 1/2 to be detected, minimum detection frame size also reducing in proportion, when not affecting testing result, decreasing the processing time of image, improve the real-time of algorithm; Timing signal reduction detection block size, namely expands and is twice.
Step 5, carries out pre-determined bit to pedestrian in the area-of-interest after piecemeal.According to the attitude information of pedestrian side under the special scenes that pedestrian crosses the street, training in advance is applicable to 12 grades of fisrt feature sorters that this scene pedestrian identifies, when locating pedestrian according to the pedestrian region, fisrt feature sorter location of training in advance, if no-fix to pedestrian region, then returns step one.This implementation step specifically comprises:
Rectangular characteristic template is adopted to travel through image, according to the fisrt feature sorter trained, pre-determined bit pedestrian region in interesting image regions.For making the fisrt feature sorter of side pedestrian, use Haar feature (i.e. rectangular characteristic template), adopt AdaBoost algorithm to a large amount of side pedestrian's samples and the training of non-pedestrian background sample, this fisrt feature sorter trains out on the basis of high number of row people sample and non-pedestrian background sample data, has stronger universality.
AdaBoost algorithm can extract the different feature of most in pedestrian's image from a large amount of pedestrian's sample gray level images, constructs one and has highly accurate pedestrian's fisrt feature sorter; Then, use the pedestrian Haar feature recorded in this fisrt feature sorter to carry out detection and positioning to pedestrian detection area-of-interest, thus obtain the pedestrian region of mating most.Utilize fisrt feature sorter and rectangular characteristic to carry out the location in pedestrian region, its judgment formula is
Wherein, H (x) represents the expression formula of one-level strong classifier, and T represents the Weak Classifier number that one-level strong classifier comprises, h
tx () represents the expression formula of t Weak Classifier, α
trepresent the weight of t Weak Classifier.
Calculated input picture region by the cascade fisrt feature sorter of 12 grades, if every one-level all exports H (x)=1, then representative navigates to pedestrian region, if there is wherein one-level to export H (x)=0, then represents and pedestrian region do not detected.
Step 6, judges the pedestrian region of pre-determined bit.There is more flase drop in the pedestrian region of extracting in step 4, by carrying out HOG (histograms of oriented gradients) feature extraction to these regions, second feature sorter that is overall according to the pedestrian trained and local judges the feature extracted, remove flase drop region, thus accurately locate pedestrian region.If pedestrian do not detected, then return step one.This implementation step specifically comprises:
1) in order to better carry out feature extraction to pre-determined bit pedestrian region to be checked, all pre-determined bit pedestrian regions to be checked can be expanded or are contracted to formed objects, such as the pedestrian's area size after all pre-determined bit is set to 64*128 fixed size, the detection block of 64*128 size comprises the block of 105 16*16 sizes, block and interblock have overlapping, the each piece of feature being expressed as 36 dimensions, the whole detection block proper vectors of 3780 dimensions describe, and calculate its 3780 dimensional feature vector.
2) according to the second feature sorter trained, the feature of each extracted region to be checked is judged, determines pedestrian region.For making the second feature sorter of side pedestrian, HOG feature extraction is carried out to a large amount of side pedestrian's samples and non-pedestrian background sample, adopt SVM algorithm, libsvm training tool is used to obtain SVM (Support Vector Machines, support vector machine) weights and threshold, i.e. the second feature sorter of side pedestrian.Second feature sorter provides weight w and threshold value b.Utilize second feature sorter and the required 3780 dimension regions of HOG feature x to pre-determined bit to judge, detecting pedestrian, its judgment formula is
Wherein, (w, b) represents the weight that second feature sorter provides and threshold value, and x represents the proper vector of input, and n is the support vector number of second feature sorter, (x
i, y
i) represent i-th support vector, λ
ifor Lagrange coefficient.
If f (x)=1, then represent and be detected as pedestrian.
3) the present invention also takes the method that local feature judges pedestrian, pedestrian is divided into head, above the waist, foot areas, respectively these three regions are combined and specify the training of people from partial row sample, obtain this trizonal sorter respectively, pre-determined bit pedestrian region is judged.
Step 7, follows the tracks of the pedestrian detected.After step 6 accurately orients pedestrian region, carrying out following the tracks of to pedestrian can the real-time of boosting algorithm greatly, is also that of step 3 section thinking supplements simultaneously.Such as the pedestrian region of the previous frame image detected is made the pedestrian's area-of-interest as this two field picture after suitably expanding, re-use step 4 method and detect.
Also can carry out integrogram calculating to image when step 5 adopts rectangular characteristic template to travel through image, integrogram is calculated as follows described in formula:
Therefore namely the traversal of rectangular characteristic template is carry out on obtained integral image, on integral image, the gray-scale value of every bit equals in original image with the gray scale summation that initial point and this point are the rectangular area to angle point, obtain each rectangular characteristic value thus, namely the eigenwert of rectangular characteristic template refers to rectangle the first half of rectangular characteristic template position and the gray scale difference of the latter half.
The present invention is by adopting the pedestrian on pedestrian detection of classifier road, and the personal feature between fuzzy pedestrian, decreases the impact of individuality difference on testing result, decrease the impact of illumination on testing result simultaneously, improve the detection efficiency of pedestrian.Simultaneously by adopting the pedestrian of crossing road on the pedestrian's attitude detection of classifier road of side, for specific application scenarios, decreasing the non-impact of crossing road pedestrian and differentiating of different attitude, improve algorithm and differentiate row human effectiveness and real-time.
Above content is in conjunction with concrete embodiment further description made for the present invention, can not assert that specific embodiment of the invention is confined to these explanations.For general technical staff of the technical field of the invention, without departing from the inventive concept of the premise, some simple deduction or replace can also be made, all should be considered as belonging to protection scope of the present invention.
Claims (7)
1. prevent a pedestrian detection method for pedestrian impact, it is characterized in that, it includes following steps:
(1) gather vehicle front image and carry out Image semantic classification;
(2) area-of-interest of pedestrian is extracted, this step is that the positional information that may collide based on pedestrian in the realtime graphic gathered is estimated, can not occur that the image that pedestrian and pedestrian are not in hazardous location is not counted in detection to part, only in more among a small circle, search for pedestrian, reduce image procossing area, to set the area-of-interest of pedestrian;
(3) piecemeal is carried out to image, reset area-of-interest;
(4) carry out one times reducing to the region to be detected of each two field picture, minimum detection frame size also reducing in proportion, reduces the processing time of image, improves algorithm real-time;
(5) in the area-of-interest after piecemeal, pre-determined bit is carried out to pedestrian;
(6) the pedestrian region of pre-determined bit is judged, specifically comprises:
Expanded or be contracted to formed objects in all pre-determined bit pedestrian regions to be checked, onesize detection block comprises multiple onesize block, and block and interblock have overlapping; According to the second feature sorter trained, the feature of each extracted region to be checked is judged; And take local feature to judge the method for pedestrian, pedestrian is divided into head, above the waist, foot areas, combines these three regions respectively and specifies the training of people from partial row sample, obtain this trizonal sorter respectively, pre-determined bit pedestrian region is judged, accurately to locate pedestrian region;
(7) pedestrian detected is followed the tracks of.
2. the pedestrian detection method of prevention pedestrian impact as claimed in claim 1, it is characterized in that, in described step (1), by being arranged on the camera collection front realtime graphic on automobile, described Image semantic classification comprises and removes serrating operation and histogram equalization operation to image.
3. the pedestrian detection method of prevention pedestrian impact as claimed in claim 1, it is characterized in that, described step (3) is that image is divided into equal some pieces, has overlapping region between block and block, wherein one piece of region is only processed to each frame input picture, each block region of circular treatment.
4. the pedestrian detection method of prevention pedestrian impact as claimed in claim 1, it is characterized in that, fisrt feature sorter pre-determined bit pedestrian region in pedestrian's area-of-interest is utilized in described step (5), described fisrt feature sorter is the feature classifiers of the attitude information of pedestrian side under the special scenes that crosses the street of the applicable pedestrian of training in advance, and described second feature sorter is the feature classifiers of the overall and local of the pedestrian of training in advance.
5. the pedestrian detection method of prevention pedestrian impact as claimed in claim 4, it is characterized in that, described fisrt feature sorter is rectangular characteristic sorter, and described second feature sorter is HOG feature classifiers.
6. the pedestrian detection method of prevention pedestrian impact as claimed in claim 5, is characterized in that, adopt AdaBoost algorithm in described step (5), adopt SVM algorithm in described step (6).
7. the pedestrian detection method of prevention pedestrian impact as claimed in claim 1, is characterized in that, carry out integrogram calculating in described step (5) to image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201110116569.0A CN102768726B (en) | 2011-05-06 | 2011-05-06 | Pedestrian detection method for preventing pedestrian collision |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201110116569.0A CN102768726B (en) | 2011-05-06 | 2011-05-06 | Pedestrian detection method for preventing pedestrian collision |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102768726A CN102768726A (en) | 2012-11-07 |
| CN102768726B true CN102768726B (en) | 2015-03-11 |
Family
ID=47096120
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201110116569.0A Active CN102768726B (en) | 2011-05-06 | 2011-05-06 | Pedestrian detection method for preventing pedestrian collision |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102768726B (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN103400113B (en) * | 2013-07-10 | 2016-08-24 | 重庆大学 | Freeway tunnel pedestrian detection method based on image procossing |
| KR101498114B1 (en) * | 2013-11-28 | 2015-03-05 | 현대모비스 주식회사 | Device and method for detecting pedestrains |
| 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 |
| CN105261017B (en) * | 2015-10-14 | 2018-12-11 | 长春工业大学 | The method that image segmentation based on road surface constraint extracts pedestrian's area-of-interest |
| CN106372666B (en) * | 2016-08-31 | 2019-07-19 | 同观科技（深圳）有限公司 | A kind of target identification method and device |
| CN107341441B (en) * | 2017-05-16 | 2020-05-08 | 开易（北京）科技有限公司 | Pedestrian collision warning system and method applied to advanced driving assistance system |
| CN108764110B (en) * | 2018-05-23 | 2021-03-23 | 大连民族大学 | Recursive false detection verification method, system and equipment based on HOG characteristic pedestrian detector |
| CN108830210B (en) * | 2018-06-11 | 2021-04-20 | 广东美的制冷设备有限公司 | Human body detection method and device based on image |
| CN114627651B (en) * | 2022-02-23 | 2023-07-25 | 深圳市锐明技术股份有限公司 | Pedestrian protection early warning method and device, electronic equipment and readable storage medium |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101556695A (en) * | 2009-05-15 | 2009-10-14 | 广东工业大学 | Image matching method |
-
2011
- 2011-05-06 CN CN201110116569.0A patent/CN102768726B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
Non-Patent Citations (3)
| Title |
|---|
| 《基于HOG的行人快速检测》;叶林 等;《计算机工程》;20101120;第3、4章 * |
| 《基于单目视觉的车辆前方行人检测技术研究》;郭烈;《吉林大学博士学位论文》;20070915;第三章，5.4节，2.1节 * |
| 《基于单目视觉的运动行人检测算法的研究与实现》;吴栓栓;《东北大学硕士学位论文》;20080601;42页第2段 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102768726A (en) | 2012-11-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102768726B (en) | Pedestrian detection method for preventing pedestrian collision | |
| Adu-Gyamfi et al. | Automated vehicle recognition with deep convolutional neural networks | |
| CN102542289B (en) | Pedestrian volume statistical method based on plurality of Gaussian counting models | |
| Li et al. | Estimating the number of people in crowded scenes by mid based foreground segmentation and head-shoulder detection | |
| Keller et al. | The benefits of dense stereo for pedestrian detection | |
| Chiverton | Helmet presence classification with motorcycle detection and tracking | |
| CN102289660B (en) | Method for detecting illegal driving behavior based on hand gesture tracking | |
| CN106682586A (en) | Method for real-time lane line detection based on vision under complex lighting conditions | |
| CN101980245B (en) | Adaptive template matching-based passenger flow statistical method | |
| CN103310194A (en) | Method for detecting head and shoulders of pedestrian in video based on overhead pixel gradient direction | |
| CN103310444B (en) | A kind of method of the monitoring people counting based on overhead camera head | |
| Li et al. | Robust people counting in video surveillance: Dataset and system | |
| CN103942577A (en) | Identity identification method based on self-established sample library and composite characters in video monitoring | |
| CN105303191A (en) | Method and apparatus for counting pedestrians in foresight monitoring scene | |
| CN105404857A (en) | Infrared-based night intelligent vehicle front pedestrian detection method | |
| CN104657724A (en) | Method for detecting pedestrians in traffic videos | |
| Møgelmose et al. | Learning to detect traffic signs: Comparative evaluation of synthetic and real-world datasets | |
| CN104134078A (en) | Automatic selection method for classifiers in people flow counting system | |
| Shirazi et al. | Contextual combination of appearance and motion for intersection videos with vehicles and pedestrians | |
| Gupta et al. | A framework for traffic light detection and recognition using deep learning and grassmann manifolds | |
| Wang et al. | Pedestrian abnormal event detection based on multi-feature fusion in traffic video | |
| Sotelo et al. | Pedestrian detection using SVM and multi-feature combination | |
| Arthi et al. | Object detection of autonomous vehicles under adverse weather conditions | |
| Tamersoy et al. | Robust vehicle detection for tracking in highway surveillance videos using unsupervised learning | |
| Cosma et al. | Part-based pedestrian detection using HoG features and vertical symmetry |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant |