CN106446789A - Pedestrian real-time detection method based on binocular vision - Google Patents
Pedestrian real-time detection method based on binocular vision Download PDFInfo
- Publication number
- CN106446789A CN106446789A CN201610778012.6A CN201610778012A CN106446789A CN 106446789 A CN106446789 A CN 106446789A CN 201610778012 A CN201610778012 A CN 201610778012A CN 106446789 A CN106446789 A CN 106446789A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- window
- feature
- real
- sample
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
Landscapes
- Engineering & Computer Science (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a pedestrian real-time detection method based on binocular vision, and the method integrates the features of eight gradient directions and the Haar-like features of RGB color channels to serve as the features, and can achieve a better detection result. Through employing the parallax information, the method reduces to-be-detected regions, is lower in complexity, is higher in speed, and can meet the real-time and accurate detection requirements. Furthermore, in order to optimize a detection result, the method comprises a training step: extracting multi-scale window features of a sample, and training multi-scale classifiers; a detection step: carrying out the feature extraction of a non-background region through employing a multi-scale window. The method provided by the invention can detect pedestrians effectively in real time, and can obtain a more precise detection result.
Description
Technical field
The present invention relates to method of video image processing, particularly to pedestrian detection technology.
Background technology
Pedestrian detection is exactly to detect whether comprise pedestrian in video or picture and mark the tram of pedestrian, and it is machine
One important branch of visual field.
The method of existing pedestrian detection is very many, but mainly uses two kinds of characteristics of image：Movable information and shape.Based on fortune
The detection method of dynamic information characteristics needs the preconditioning techniques such as background extracting and image segmentation, and the detection side based on shape facility
Method does not need to use Preprocessing Algorithm.
Detection method based on shape facility is divided into global characteristics method and local characteristic method according to the extracting method of feature.
The difference of global characteristics and local feature is that global characteristics to extract feature from whole image, and local feature is from image
Regional area extracting feature.The classical example of global characteristics is independent principal component analysis PCA, and its shortcoming is to object
Appearance, posture and light sensitive, and local feature is due to extracting feature from the local of image, to the appearance of object, posture and
Illumination is more insensitive.Typical local feature has wavelet coefficient, gradient direction and local covariance etc..Local feature can enter again
One step is divided into whole detection and location detection, and the testing result of location detection is by the detection at position by another one grader
Result is combined into final pedestrian detection result.Advantage using location detection method is that it can be tackled well due to pedestrian's limb
Body movement leads to the problem that the change of pedestrian's appearance brings, and shortcoming is that it makes whole detection process become more complicated.
It is the most frequently used and effective method of current pedestrian detection based on the method for statistical learning, the method passes through a large amount of training
Sample builds pedestrian detection grader.The feature extracted typically has the gray scale of target, edge, texture, shape, histogram of gradients etc.
Information, grader includes neutral net, SVM, Adaboost etc..There is following difficult point in the method：The attitude of pedestrian, dress ornament are respectively not
Identical, not compact, grader the performance of distribution in feature space for the feature extracted by training sample affected larger, from
Negative sample during line training cannot cover the situation of all true application scenarios.
Content of the invention
The technical problem to be solved is to provide a kind of fast and effectively to be examined based on the pedestrian of binocular vision in real time
Survey method.
The present invention be employed technical scheme comprise that by solving above-mentioned technical problem, the pedestrian's real-time detection based on binocular vision
Method, comprises the following steps：
1) training step：Extract sample characteristics input grader to be trained；
2) detecting step：
2-1) utilize binocular camera to gather video to be detected, calculate the disparity map of each two field picture in video to be detected；
2-2) only the region of threshold value is more than or equal to as non-background area to parallax value in testing image；
2-3) the feature input grader of non-background area is carried out pedestrian detection.
Further, the present invention is used 8 gradient direction features to be levied with the class Lis Hartel of RGB color passage and combines as special
Levy, preferable Detection results can be reached.
Further, for optimizing detection effect, in training step, sample is carried out extracting multiple dimensioned window feature,
Train multiple yardstick graders；In detecting step, non-background area is carried out carry out feature extraction using multiple dimensioned window.
Further, in order to accelerate feature extraction speed, in training step, first the window using a size of 4 × 4 is sliding on sample
Move and complete feature extraction, afterwards by sample being zoomed in and out with the window reusing a size of 4 × 4 in the sample being scaled
Upper slip complete feature extraction obtain multiple dimensioned under window feature extract；In detecting step, first use a size of 4 × 4 window
Mouth slides on image and completes feature extraction；After the completion of non-background area judges, slided on image with each size window, sentence
Proportion shared by region to be detected in disconnected window, is then considered as not comprising pedestrian when proportion is less than default proportion, does not carry out point
Class device judges, when proportion is more than or equal to default proportion, then selects current window position institute in the feature under 4 × 4 sizes
Corresponding feature simultaneously determines whether to comprise pedestrian with grader.
The invention has the beneficial effects as follows, the method based on binocular vision can detect pedestrian effectively in real time, and can reach
One more accurate testing result；Decrease region to be detected using parallax information, complexity is lower, speed faster, Neng Gouman
Requirement is accurately detected when full.
Brief description
Fig. 1：The pedestrian detection schematic flow sheet of the present invention.
Specific embodiment
The present invention can be divided into training and two stages of detection, as shown in Figure 1.
Comprise pedestrian under collection different scenes first and do not comprise the picture of pedestrian as sample, be trained being detected
Parameter, the training stage specifically can be divided into following four step：
Step one：It is 64 × 128 5000 positive samples with size and 5000 negative samples are trained；
Step 2：For ultimate unit, feature extraction is carried out with one 4 × 4 not overlapping and adjacent block to samples pictures, uses
Feature be 8 gradient direction features, the class Lis Hartel of RGB color passage levies.For 8 gradient direction features, image is any
Some computing formula are as follows：
Wherein, θiFor the gradient direction quantifying, its value is i, and G (x, y) is gradient magnitude, and R (x, y) is the arc of gradient direction
Angle value, Θ (x, y) is the quantized value of gradient direction,Value for corresponding gradient direction.The calculating of gradient is calculated using [- 10 1]
Son, calculates gradient respectively on tri- passages of RGB, take gradient magnitude maximum for final gradient magnitude and gradient direction.System
Meter 4 × 4 pieces in all 8 different directions value and and average to obtain the corresponding eigenvalue of all directions, finally connect
Obtain 8 gradient direction features F in 4 × 4 piecesg′.
It is to calculate spy respectively in 3 passages in the same way that the class Lis Hartel of RGB color passage levies computational methods
Levy., in calculating 4 × 4 pieces, each 2 × 2 not overlapping and the sum of adjacent block, can obtain 4 values, be set to B taking R passage as a example0,B1,
B2,B3(sequentially for from left to right, from top to bottom), feature calculation formula is as follows：
Ll=B0+B1+B2+B3(4)
Fh=[ll lh hl hh] (8)
The class Lis Hartel of RGB color passage levies Fh' it is that the feature string connection on three passages is combined.
Eventually for the feature trained and detect it is
F=[Fq′Fh′] (9)
Step 2：Extract difficult example using the detection being obtained in negative sample, then gained hardly possible example is added negative sample simultaneously
Therefrom take out 10000 negative samples at random as the negative sample of second training, 10000 negative samples are by this and 5000 before
Training obtains grader together for positive sample training.
Step 3：The size scaling of sample is 72 × 144,76 × 152, then is directed to sample by step one, two, three acquisitions
This size of 72 × 144 and 76 × 152 detection parameter.Afterwards, retraining is carried out to the negative sample being detected as positive sample, that is,
Extract difficult example retraining, so that it may carry out pedestrian detection after multiple training obtains detection parameter.
Detection process is detected by multiple dimensioned multiwindow, and detection part is broadly divided into following six step：
Step one：Initially calculate eigenvalue, the computational methods of feature and the instruction at each 4x4 window of entire image
The computational methods practicing stage etch one are consistent.
Step 2：Calculate the parallax information of image by seeking the block matching algorithm of the absolute value sum of corresponding blocks difference,
And image is divided into by background area and non-background area according to parallax value size.
Step 3：Window with a size of 64 × 128 slides on image, and ranks stepping is 8, often slides into certain
Position, first judges the proportion shared by non-background parallax in window, if proportion is little, is considered as not comprising pedestrian, otherwise just carries
Take the feature of correspondence position in image, and determine whether pedestrian using corresponding detection parameter.The sentencing of proportion size
Open close mistake and default proportion are compared to realize.
Step 4：Window size is changed to 72 × 144,76 × 152, repeat step three.
Step 5：Image is carried out up-sampling with 1.33 ratios and down-sampling is multiple, the picture size after down-sampling is necessary
More than 76 × 152, the image after up-sampling is defined according to actual needs.Step is repeated after image after being sampled every time
Rapid one, two, three, four.
Step 6：The window that the judgement obtaining is pedestrian carries out obtaining final result after non-maxima suppression.
Claims (4)
1. the pedestrian's real-time detection method based on binocular vision is it is characterised in that comprise the following steps：
1) training step：Extract sample characteristics input grader to be trained；
2) detecting step：
2-1) utilize binocular camera to gather video to be detected, calculate the disparity map of each two field picture in video to be detected；
2-2) only the region of threshold value is more than or equal to as non-background area to parallax value in image；
2-3) the feature input grader of non-background area is carried out pedestrian detection.
2. the pedestrian's real-time detection method based on binocular vision as claimed in claim 1 is it is characterised in that use 8 gradient sides
Class Lis Hartel to characteristic binding RGB color passage is levied as feature.
3. as claimed in claim 1 or 2 the pedestrian's real-time detection method based on binocular vision it is characterised in that in training step
In, extract the multiple dimensioned window feature of sample, train multiple yardstick graders；
In detecting step, Analysis On Multi-scale Features are extracted to non-background area and detects.
4. as claimed in claim 3 the pedestrian's real-time detection method based on binocular vision it is characterised in that in training step,
First slided on sample using a size of 4 × 4 window and complete feature extraction, reuse chi by zooming in and out to sample afterwards
Very little be 4 × 4 window slide on the sample being scaled complete feature extraction obtain multiple dimensioned under window feature extract；
In detecting step, first use a size of 4 × 4 window to slide on image and complete feature extraction；Sentence in non-background area
After the completion of disconnected, slided on image with each size window, judge the proportion shared by region to be detected in window, when proportion is little
Then it is considered as not comprising pedestrian in default proportion, does not carry out grader judgement, when proportion is more than or equal in default proportion, then exist
Select the feature corresponding to current window position in feature under 4 × 4 sizes and determine whether to comprise pedestrian with grader.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610778012.6A CN106446789A (en) | 2016-08-30 | 2016-08-30 | Pedestrian real-time detection method based on binocular vision |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610778012.6A CN106446789A (en) | 2016-08-30 | 2016-08-30 | Pedestrian real-time detection method based on binocular vision |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106446789A true CN106446789A (en) | 2017-02-22 |
Family
ID=58091759
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610778012.6A Pending CN106446789A (en) | 2016-08-30 | 2016-08-30 | Pedestrian real-time detection method based on binocular vision |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106446789A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107292927A (en) * | 2017-06-13 | 2017-10-24 | 厦门大学 | A kind of symmetric motion platform's position and pose measuring method based on binocular vision |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104036284A (en) * | 2014-05-12 | 2014-09-10 | 沈阳航空航天大学 | Adaboost algorithm based multi-scale pedestrian detection method |
| CN104504688A (en) * | 2014-12-10 | 2015-04-08 | 上海大学 | Method and system based on binocular stereoscopic vision for passenger flow density estimation |
| CN104902258A (en) * | 2015-06-09 | 2015-09-09 | 公安部第三研究所 | Multi-scene pedestrian volume counting method and system based on stereoscopic vision and binocular camera |
| US20160154993A1 (en) * | 2014-12-01 | 2016-06-02 | Modiface Inc. | Automatic segmentation of hair in images |
| CN105760858A (en) * | 2016-03-21 | 2016-07-13 | 东南大学 | Pedestrian detection method and apparatus based on Haar-like intermediate layer filtering features |
-
2016
- 2016-08-30 CN CN201610778012.6A patent/CN106446789A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104036284A (en) * | 2014-05-12 | 2014-09-10 | 沈阳航空航天大学 | Adaboost algorithm based multi-scale pedestrian detection method |
| US20160154993A1 (en) * | 2014-12-01 | 2016-06-02 | Modiface Inc. | Automatic segmentation of hair in images |
| CN104504688A (en) * | 2014-12-10 | 2015-04-08 | 上海大学 | Method and system based on binocular stereoscopic vision for passenger flow density estimation |
| CN104902258A (en) * | 2015-06-09 | 2015-09-09 | 公安部第三研究所 | Multi-scene pedestrian volume counting method and system based on stereoscopic vision and binocular camera |
| CN105760858A (en) * | 2016-03-21 | 2016-07-13 | 东南大学 | Pedestrian detection method and apparatus based on Haar-like intermediate layer filtering features |
Non-Patent Citations (2)
| Title |
|---|
| XIAOHUI LIU等: "A pedestrian detection system based on binocular stereo", 《 2012 INTERNATIONAL CONFERENCE ON WIRELESS COMMUNICATIONS AND SIGNAL PROCESSING (WCSP)》 * |
| 李梦涵等: "多尺度级联行人检测算法的研究与实现", 《计算机技术与发展》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107292927A (en) * | 2017-06-13 | 2017-10-24 | 厦门大学 | A kind of symmetric motion platform's position and pose measuring method based on binocular vision |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105512683B (en) | Object localization method and device based on convolutional neural networks | |
| CN110119728A (en) | Remote sensing images cloud detection method of optic based on Multiscale Fusion semantic segmentation network | |
| CN103400156B (en) | Based on the High Resolution SAR image Ship Detection of CFAR and rarefaction representation | |
| CN102722891B (en) | Method for detecting image significance | |
| CN104850850B (en) | A kind of binocular stereo vision image characteristic extracting method of combination shape and color | |
| CN109284669A (en) | Pedestrian detection method based on Mask RCNN | |
| CN107273832B (en) | License plate recognition method and system based on integral channel characteristics and convolutional neural network | |
| CN109711288A (en) | Remote sensing ship detecting method based on feature pyramid and distance restraint FCN | |
| CN104778721A (en) | Distance measuring method of significant target in binocular image | |
| CN109446922B (en) | Real-time robust face detection method | |
| CN106557740B (en) | The recognition methods of oil depot target in a kind of remote sensing images | |
| CN106446890B (en) | A kind of candidate region extracting method based on window marking and super-pixel segmentation | |
| CN107066916A (en) | Scene Semantics dividing method based on deconvolution neutral net | |
| CN105069774B (en) | The Target Segmentation method of optimization is cut based on multi-instance learning and figure | |
| CN104537689B (en) | Method for tracking target based on local contrast conspicuousness union feature | |
| CN110298297A (en) | Flame identification method and device | |
| CN105678318B (en) | The matching process and device of traffic sign | |
| CN104517095A (en) | Head division method based on depth image | |
| CN108268865A (en) | Licence plate recognition method and system under a kind of natural scene based on concatenated convolutional network | |
| Galsgaard et al. | Circular hough transform and local circularity measure for weight estimation of a graph-cut based wood stack measurement | |
| CN114926747A (en) | Remote sensing image directional target detection method based on multi-feature aggregation and interaction | |
| CN108073940A (en) | A kind of method of 3D object instance object detections in unstructured moving grids | |
| CN106529441A (en) | Fuzzy boundary fragmentation-based depth motion map human body action recognition method | |
| CN105354547A (en) | Pedestrian detection method in combination of texture and color features | |
| Dousai et al. | Detecting humans in search and rescue operations based on ensemble learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170222 |
|
| RJ01 | Rejection of invention patent application after publication |