CN108288041A - A kind of preprocess method of pedestrian target false retrieval removal - Google Patents
A kind of preprocess method of pedestrian target false retrieval removal Download PDFInfo
- Publication number
- CN108288041A CN108288041A CN201810078599.9A CN201810078599A CN108288041A CN 108288041 A CN108288041 A CN 108288041A CN 201810078599 A CN201810078599 A CN 201810078599A CN 108288041 A CN108288041 A CN 108288041A
- Authority
- CN
- China
- Prior art keywords
- image
- pedestrian
- normalization
- false retrieval
- width
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/30—Noise filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
Abstract
A kind of preprocess method of pedestrian target false retrieval removal, belongs to pedestrian detection preprocessing technical field, the false retrieval phenomenon common for solving pedestrian detection result in image, and technical essential is：S1. normalization width processing；S2. histogram specification equalization processing；S3. Edge contrast, effect are：The redundancy pedestrian's detection block for having effectively removed the appearance during pedestrian detection promotes the validity of follow-up pedestrian's track algorithm application, false retrieval target is avoided to calculate consumption to the additional judgement that pedestrian detecting system is brought.
Description
Technical field
The invention belongs to pedestrian detection preprocessing technical fields, specifically a kind of to reduce the wrong pedestrian detected in image
The method of destination number.
Background technology
It is also growing to intelligence machine Man's Demands with the rapid development in robot application field.In recent years, row
People's detection technique has become the important component of one of hot research direction and computer vision field.Common pedestrian
Detection method is simulation human eye, realizes that machine carries out Image Acquisition to ambient enviroment or things, then in the image of acquisition
Appearance is analyzed and is identified, perception of the machine to three-dimensional world is completed.
The detection thought of the pedestrian detector of mainstream mainly divides two major classes at present：Machine learning detection method and deep learning
Detection method.The common ground of these two kinds of methods, with Vision information processing technology, is realized using picture as primary information resource
Circle choosing to pedestrian target.
During pedestrian's detection process, image to be detected can be because of different scenes difference, uneven light, intensive shelter
Etc. factors and cause the complexity of detection information to be promoted, it is easy to so that the characteristic information of pedestrian the possibility of flase drop is occurred, to
Reduce the identification accuracy of detecting system.In general, the false retrievals phenomenon such as the false retrieval of pedestrian detection frame, redundancy also becomes autonomous
It drives, the vehicle-mounted problem for assisting not avoiding in the applications such as driving, robot vision and pedestrian's avoidance.
For environment such as actual urban roads, due to including building, road markings, light show window, vehicle carrier etc.
Information, any include the image of pedestrian is complex background condition, and laboratory or that ecotopia of film studio is not present.Institute
With pedestrian's false retrieval removal technology for this scene is very necessary.
In general, by multi-detector fusion method, based on partial model detection method, based on the differences such as assemblage characteristic algorithm
Targeting algorithms ensure the accuracy of system, but as above-mentioned targetedly algorithm is only capable of using special screne, do not have
The universality of application.Application publication number is that the Chinese invention patent application of CN106650667A discloses one kind based on supporting vector
The pedestrian detection method and system of machine, refer to support vector machines cannot well controlling feature weight distribution and reduce pedestrian's mesh
Accuracy in detection problem is marked, the genetic algorithm of simulated annealing acceptance criterion is proposed, considers in image and contacted between feature, be
The feature of all dimensions formulates weights, to improve the accuracy of pedestrian detection；According to application publication number CN105913003A's
A kind of pedestrian detection method for multiple features multi-model that Chinese invention patent application is recorded, then be to analyze a variety of pedestrian detections
The advantage and disadvantage of method propose that a kind of pedestrian detection method of multiple features multi-model, the content of comprehensive a variety of pedestrian detectors are realized
The purpose of compatible high detection rate and low rate of false alarm simultaneously.
Invention content
Emphasis of the present invention is to pre-processing monitoring image using the combination for having image processing means, to realize removal
The purpose of lengthy and jumbled pedestrian detection information.According to contour feature machine learning " flight data recorder " internal feature parameter, false retrieval mesh is determined
Mark removes pretreated specific processing sequence and the suitable parameters of each processing method.
The width that the invention is only applicable to input image resolution is more than or equal to 75 pixels, and less than or equal to 1080 pixels
Image, i.e. picture traverse require simultaneously in [75,1080] closed interval, and the height of input picture is less than or equal to 1920 pixels.
Since the equipment of acquisition image is different, various noises and distortion included in original image to be detected, by
This generates false retrieval target.Currently, the pedestrian detector of view-based access control model image is realized by calculating the characteristics of image of image to be detected
Correct classification to objects in images.And the pedestrian detector based on contour feature is then more sensitive to shape feature in image,
Except factors such as external environments, during actual pedestrian detection, react the texture information of details reduces to a certain extent
The accuracy of detection of such pedestrian detector.So devising a kind of pedestrian for false retrieval phenomenon common in pedestrian detection result
Target false retrieval removes preprocess method, and this method removes the redundancy in original image to be detected first with normalization width processing
High fdrequency component, filter out disturbance factor；Then to treated, image introduces histogram specification equilibrium treatment and Edge contrast
Make image outline information sharpening, highlight the marginal information in image, increases the identifiability of shape feature.
Furtherly, this method is by three processing of normalization width, histogram specification equilibrium treatment and Edge contrast elder generations
Orderly processing links composition afterwards, and processing sequence can not be changed, three processing links means are organically cascaded, and are formed
One effective image preprocessing chain combination.Meanwhile each link suffers from stringent inherent parameter regulation, it is specific as follows.
(1) redundancy removes
1) normalization width processing
First link of the width processing as entire preprocess method is normalized, all inputs are all subjected to wide constraint.
Its essence is image equal proportion to be zoomed to specified pixel width, the width using the conventional means in digital image processing techniques
Degree is derived from by feature pyramid in the machine learning based on contour feature.The numerical value is closer to the pyramidal characterization of feature
Parameter, required calculation amount is not only smaller, but also makes the testing result of the pedestrian detector based on contour feature also more smart
Really.The final numerical value is confirmed as 75,150 and 225.
If the height of original image to be detected, wide respectively H0And W0, then image scaled r be：
It is another that set normalization width treated and is high, wide be respectively H1And W1, then treated, and image resolution sizes calculate such as
Under：
It fully enters image after normalization width processing by the first link and fixed width is zoomed to by equal proportion
Resolution dimensions.
(2) contour feature enhances
1) histogram specification equilibrium treatment
Second link of pedestrian target false retrieval removal preprocess method is carried out to the image that previous link is disposed
Histogram specification equilibrium treatment.The enhancing, which handles most important purpose, to be detached bright with dark things, is based on to be promoted
The pedestrian detector of contour feature distinguishes the probability of target.
There are two the multi-modal Gauss methods of Gaussian function to carry out regulation equilibrium treatment using tool by the present invention.First height
The mean value and variance of this function be respectively：0.136 and 0.045；The mean value 0.764 and 0.045 of second Gaussian function.
By the second link, image can carry out coloration enhancing in prescribed limit.
2) Edge contrast
Pedestrian target false retrieval removes the third link of preprocess method, is carried out to the image that previous link is disposed sharp
Change is handled.Its purpose is to effectively improve object edge profile, to reinforce the differentiation degree of front and back scenery in scene, with this
Improve the detectability of the pedestrian detector based on contour feature.It sharpens mould：
After the orderly link that links up by three above, image has obtained texture etc. through specification processing and has obtained appropriately
Enhancing, is handled to be very suitable for being further fed into pedestrian detector.
Advantageous effect：
The present invention is exactly to make image after the pretreatment of the above method, and the wrong pedestrian's destination number detected can substantially drop
Low, the pedestrian target that as the input preprocessing means of popular pedestrian detector, can be suitable for street scene is wrong
Examine elimination demand.Can effectively by pedestrian detection since arithmetic accuracy is not high, input picture scene is excessively complicated and not phase
The removal of redundancy, false retrieval pedestrian detection frame caused by the factors such as dry characteristic information, to promote the detection of pedestrian detecting system
Order of accuarcy；Preprocess method is removed by false retrieval pedestrian target, enables a system to avoid in practical applications because of taking lens
The factors such as angle, photographed scene uneven illumination be even lead to the detection that complex background image to be detected is likely to occur in detection calculations
Error situation.The present invention has effectively removed redundancy pedestrian's detection block of the appearance during pedestrian detection, promoted follow-up pedestrian with
The validity of track algorithm application avoids false retrieval target from calculating consumption to the additional judgement that pedestrian detecting system is brought.
The present invention using normalization width processing, improves the picture quality of original image to be detected, to return first
One changes width processing as the main means for filtering out image medium-high frequency redundancy；Histogram specification equilibrium treatment is for increasing
Picture contrast expands quantized interval, keeps each interregional boundary apparent, and Edge contrast is then to improve interregional contrast, makes side
Boundary is more obvious, the contour feature of prominent image, histogram specification equilibrium treatment and Edge contrast jointly to image implementation at
Reason, makes image edge shape feature sharpening, highlights the information content that the features of shape in image reduces redundancy, increases recognizable
Property.
The present invention is suitable as the preprocessor of arbitrary pedestrian detector, and processing method is simple, it can be achieved that property is strong；Normalizing
Change width, histogram specification equilibrium, sharpen three kinds of processing steps, parameter is fixed, and can be used without priori；It is used
Processing method also belong to classical image processing means, be easy to various soft hardware equipments realize.
Description of the drawings
Fig. 1 is the principle logic chart using the pedestrian detection method of pedestrian target false retrieval removal preprocess method；
Fig. 2 is to remove to detect knot obtained from preprocess method without pedestrian target false retrieval for the image to be detected chosen
Fruit；
Fig. 3 is the inspection obtained after the removal preprocess method processing of pedestrian target false retrieval for the image to be detected chosen
Survey result；
Fig. 4 is the testing result figure of embodiment 1；
Fig. 5 is the testing result figure of embodiment 2；
Fig. 6 is the testing result figure of embodiment 3.
Specific implementation mode
Present invention is further described in detail with specific implementation mode below in conjunction with the accompanying drawings：
A kind of principle logic chart of the preprocess method of pedestrian target false retrieval removal is as shown in Figure 1, the algorithm is embodied
Steps are as follows：
Present invention is further described in detail with specific implementation mode below in conjunction with the accompanying drawings：
A kind of principle logic chart of the preprocess method of pedestrian target false retrieval removal is as shown in Figure 1, the algorithm is embodied
Steps are as follows：
1st step：Obtain the wide W of image0, high H0, calculate the ratio r of picture to be detected；
2nd step：Judge picture width W0Affiliated range, implement the normalization width processing of first link；
3rd step：Work as W0∈ [75,200) when, width, high respectively W after normalization1=75, H1=r × W1, go to the 6th
Step；
4th step：Work as W0∈ [200,800) when, width, high respectively W after normalization1=150, H1=r × W1, go to the 6th
Step；
5th step：Work as W0∈ [800,1080) when, width, high respectively W after normalization1=225, H1=r × W1, go to
6 steps；
6th step：Into second link.Wide, a height of W of image after normalization1And H1, use multi-modal Gaussian function pair
Image does histogram specification equilibrium treatment；
7th step：Into third link.Image after histogram specification equilibrium treatment is wide, a height of W1And H1, utilize mould
Plate A does Edge contrast；
8th step：The image to be detected being disposed.
The above method is that a kind of input to detector pre-processes, to make pedestrian target mistake part reduced number
Method.The processing method is using three kinds of processing of normalization width, histogram specification equilibrium treatment, Edge contrast means to detection
The input picture of device is converted.Through a large amount of emulation testings, the method for the present invention can promote the reliability of pedestrian detection algorithm, greatly
Width reduces the phenomenon that detector false retrieval target pedestrian, possesses very strong practicability in directions such as feature calculation, convolutional neural networks.
The present invention " inhibits noise " first, is realized to the noise reduction of original image to be detected using normalization width processing
Reason, filters out image high frequency components；Then " enhancing image ", the image to filtering out high-frequency noise are implemented at contour feature enhancing
Reason introduces histogram specification equilibrium treatment means enlarged image dynamic range here, so that image-region is divided apparent, at sharpening
Reason means then increase interregional contrast, prominent image edge information.The processing method is using the processing of normalization width, histogram
Three kinds of regulation equilibrium treatment, Edge contrast means convert the input picture of detector.It can be seen in various documents
It arrives, the processing mode of this " first inhibiting to enhance afterwards " is more universal, still " inhibits " means innumerable with " enhancing " means, can
It is that the maximum problem of method of machine learning class and deep learning class is themselves to be exactly one " flight data recorder ", you can not do
The tendency of pre-control study, then how to carry out screening for the specific machine learning characteristic of contour feature needs a large amount of theory
Guidance and practical basis.Through emulation testing, what normal process method determined by the present invention can promote pedestrian detection algorithm can
By property, the phenomenon that detector false retrieval target pedestrian is greatly decreased, possess in directions such as feature calculation, convolutional neural networks very strong
Practicability.
Embodiment 1：
The present embodiment is directed to picture traverse W0∈ [75,200) practical pedestrian detector input picture frame, this is sent out
The bright input applied to pedestrian detector pre-processes, the testing result before processing and treated testing result comparison diagram such as Fig. 4
In (a), shown in (b), effectively remove the pedestrian target of false retrieval.
Embodiment 2：
The present embodiment is directed to picture traverse W0∈ [test by the input picture frame of practical pedestrian detector 200,800)
Card, practical image to be detected characteristic information complexity leads to the pedestrian target frame false retrieval phenomenon of output, using the inspection before and after the present invention
It surveys comparative result figure such as (a) in Fig. 5, shown in (b), successfully remove the output result of false retrieval pedestrian target.
Embodiment 3：
The present embodiment is directed to picture traverse W0[800,1080) the input picture frame of practical pedestrian detector carries out real ∈
It tests, the present invention can effectively remove the pedestrian target false retrieval phenomenon that shape can not be resisted to lead to output due to practical pedestrian to be detected, most
Eventually the front and back pedestrian detection comparative result figure of pretreatment such as (a) in Fig. 6, shown in (b).
Claims (5)
1. a kind of preprocess method of pedestrian target false retrieval removal, which is characterized in that include the following steps：
S1. normalization width processing；
S2. histogram specification equalization processing；
S3. Edge contrast.
2. the preprocess method of pedestrian target false retrieval removal as described in claim 1, which is characterized in that normalization width processing
The step of it is as follows：If the height of original image to be detected, wide respectively H0And W0, then image scaled r be：
It is another that set normalization width treated and is high, wide be respectively H1And W1, then treated, and image resolution sizes calculating is as follows：
3. the preprocess method of pedestrian target false retrieval removal as described in claim 1, which is characterized in that histogram specification is equal
The step of weighing apparatusization processing, is as follows：Using tool, there are two the multi-modal Gauss methods of Gaussian function to carry out regulation equilibrium treatment, the
The mean value and variance of one Gaussian function be respectively：0.136 and 0.045；The mean value 0.764 and 0.045 of second Gaussian function.
4. the preprocess method of pedestrian target false retrieval removal as described in claim 1, which is characterized in that the sharpening of Edge contrast
Mould is：
5. a kind of preprocess method of pedestrian target false retrieval removal, which is characterized in that include the following steps：
1st step：Obtain the wide W of image0, high H0, calculate the ratio r of picture to be detected；
2nd step：Judge picture width W0Affiliated range, implement normalization width processing；
3rd step：Work as W0∈ [75,200) when, width, high respectively W after normalization1=75, H1=r × W1, go to the 6th step；
4th step：Work as W0∈ [200,800) when, width, high respectively W after normalization1=150, H1=r × W1, go to the 6th step；
5th step：Work as W0∈ [800,1080) when, width, high respectively W after normalization1=225, H1=r × W1, go to the 6th step；
6th step：Wide, a height of W of image after normalization1And H1, it is equal that histogram specification is done to image using multi-modal Gaussian function
Weighing apparatus processing；
7th step：Image after histogram specification equilibrium treatment is wide, a height of W1And H1, Edge contrast is done using template A；
8th step：The image to be detected being disposed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810078599.9A CN108288041B (en) | 2018-01-26 | 2018-01-26 | Preprocessing method for removing false detection of pedestrian target |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810078599.9A CN108288041B (en) | 2018-01-26 | 2018-01-26 | Preprocessing method for removing false detection of pedestrian target |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108288041A true CN108288041A (en) | 2018-07-17 |
| CN108288041B CN108288041B (en) | 2021-02-02 |
Family
ID=62835982
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810078599.9A Active CN108288041B (en) | 2018-01-26 | 2018-01-26 | Preprocessing method for removing false detection of pedestrian target |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108288041B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110457823A (en) * | 2019-08-13 | 2019-11-15 | 大连民族大学 | The MLP method of super-elasticity cylindrical thin shell strong nonlinear vibration |
| CN111698459A (en) * | 2019-04-26 | 2020-09-22 | 泰州阿法光电科技有限公司 | Real-time analysis method for object parameters |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103093186A (en) * | 2011-07-28 | 2013-05-08 | 施乐公司 | Systems and methods for improving image recognition |
| CN103136530A (en) * | 2013-02-04 | 2013-06-05 | 国核自仪系统工程有限公司 | Method for automatically recognizing target images in video images under complex industrial environment |
| CN103218621A (en) * | 2013-04-21 | 2013-07-24 | 北京航空航天大学 | Identification method of multi-scale vehicles in outdoor video surveillance |
| CN103593648A (en) * | 2013-10-22 | 2014-02-19 | 上海交通大学 | Face recognition method for open environment |
| CN105335643A (en) * | 2015-10-28 | 2016-02-17 | 广东欧珀移动通信有限公司 | Processing method and processing system of files |
| CN107527333A (en) * | 2017-07-31 | 2017-12-29 | 湖北工业大学 | A kind of rapid image Enhancement Method based on gamma transformation |
-
2018
- 2018-01-26 CN CN201810078599.9A patent/CN108288041B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103093186A (en) * | 2011-07-28 | 2013-05-08 | 施乐公司 | Systems and methods for improving image recognition |
| CN103136530A (en) * | 2013-02-04 | 2013-06-05 | 国核自仪系统工程有限公司 | Method for automatically recognizing target images in video images under complex industrial environment |
| CN103218621A (en) * | 2013-04-21 | 2013-07-24 | 北京航空航天大学 | Identification method of multi-scale vehicles in outdoor video surveillance |
| CN103593648A (en) * | 2013-10-22 | 2014-02-19 | 上海交通大学 | Face recognition method for open environment |
| CN105335643A (en) * | 2015-10-28 | 2016-02-17 | 广东欧珀移动通信有限公司 | Processing method and processing system of files |
| CN107527333A (en) * | 2017-07-31 | 2017-12-29 | 湖北工业大学 | A kind of rapid image Enhancement Method based on gamma transformation |
Non-Patent Citations (1)
| Title |
|---|
| RIFAT MUEID等: "Vehicle-Pedestrian Dynamic Interaction through Tractography of Relative Movements and Articulated Pedestrian Pose Estimation", 《2016 IEEE APPLIED IMAGERY PATTERN RECOGNITION WORKSHOP (AIPR)》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111698459A (en) * | 2019-04-26 | 2020-09-22 | 泰州阿法光电科技有限公司 | Real-time analysis method for object parameters |
| CN111698459B (en) * | 2019-04-26 | 2021-07-27 | 广东邦盛北斗科技股份公司 | Real-time analysis method for object parameters |
| CN110457823A (en) * | 2019-08-13 | 2019-11-15 | 大连民族大学 | The MLP method of super-elasticity cylindrical thin shell strong nonlinear vibration |
| CN110457823B (en) * | 2019-08-13 | 2023-06-02 | 大连民族大学 | MLP method for strong nonlinear vibration of superelastic cylindrical thin shell |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108288041B (en) | 2021-02-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107330376B (en) | Lane line identification method and system | |
| Siriborvornratanakul | An automatic road distress visual inspection system using an onboard in-car camera | |
| CN107273832B (en) | License plate recognition method and system based on integral channel characteristics and convolutional neural network | |
| CN103824070B (en) | A kind of rapid pedestrian detection method based on computer vision | |
| CN105894701B (en) | The identification alarm method of transmission line of electricity external force damage prevention Large Construction vehicle | |
| CN104517104A (en) | Face recognition method and face recognition system based on monitoring scene | |
| CN106250838A (en) | vehicle identification method and system | |
| Li et al. | Road lane detection with gabor filters | |
| CN107220649A (en) | A kind of plain color cloth defects detection and sorting technique | |
| Derpanis et al. | Classification of traffic video based on a spatiotemporal orientation analysis | |
| CN105373779A (en) | Intelligent vehicle safety belt detection method and intelligent detection system | |
| CN109918971B (en) | Method and device for detecting number of people in monitoring video | |
| CN103632170A (en) | Pedestrian detection method and device based on characteristic combination | |
| Kashevnik et al. | Seat belt fastness detection based on image analysis from vehicle in-abin camera | |
| CN106548131A (en) | A kind of workmen's safety helmet real-time detection method based on pedestrian detection | |
| Yuen et al. | On looking at faces in an automobile: Issues, algorithms and evaluation on naturalistic driving dataset | |
| CN109902576B (en) | Training method and application of head and shoulder image classifier | |
| CN106127812A (en) | A kind of passenger flow statistical method of non-gate area, passenger station based on video monitoring | |
| Song et al. | Weather condition recognition based on feature extraction and K-NN | |
| CN105740751A (en) | Object detection and identification method and system | |
| Malhi et al. | Vision based intelligent traffic management system | |
| CN108460320A (en) | Based on the monitor video accident detection method for improving unit analysis | |
| US20070223785A1 (en) | Image processor and method | |
| Farag | A comprehensive real-time road-lanes tracking technique for autonomous driving | |
| CN105354547A (en) | Pedestrian detection method in combination of texture and color features |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |