CN104978570B - The detection and recognition methods of traffic sign in driving video based on incremental learning - Google Patents
The detection and recognition methods of traffic sign in driving video based on incremental learning Download PDFInfo
- Publication number
- CN104978570B CN104978570B CN201510359501.3A CN201510359501A CN104978570B CN 104978570 B CN104978570 B CN 104978570B CN 201510359501 A CN201510359501 A CN 201510359501A CN 104978570 B CN104978570 B CN 104978570B
- Authority
- CN
- China
- Prior art keywords
- traffic sign
- detection
- result
- detector
- tracking
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Abstract
The technical issues of detection and recognition methods of traffic sign in the driving video that the invention discloses a kind of based on incremental learning, detection for solving existing traffic sign and recognition methods poor robustness.Technical solution is to train Adaboost graders using converging channels feature, and detector is improved, then the tracking based on motion model is carried out using the testing result of the detector as the observation of Kalman filter, during tracking, the new increment SVM detectors of on-line training, when former Adaboost detectors lead to detection failure due to the apparent variation of mark, it is detected by the online incremental detector, and inputted testing result as the observation of Kalman filter, the target for failing to be consecutively detected is filtered.The Nearest Neighbor with Weighted Voting that the Gauss weight based on scale is finally carried out to the tracking result of the same physics traffic sign, obtains final recognition result, improves the robustness of detection and identification.
Description
Technical field
The present invention relates to a kind of detection of traffic sign and recognition methods, more particularly to a kind of row based on incremental learning
The detection and recognition methods of traffic sign in vehicle video.
Background technology
Document " AndreasDongran Liu,Mohan M.Trivedi,Traffic Sign
Detection for U.S.Roads:Remaining Challenges and a case for Tracking.IEEE
Using cascade point in Intelligent Transportation Systems Conference, pp.1394-1399.2014. "
Class device Adaboost methods are detected.This article selects integrating channel feature to carry out feature extraction to input training data, then
The Adaboost cascade classifiers of one 3 layers of training, the method that sliding window is used in combination are detected inputting video data.It is this
Method for traffic sign detection excessively relies on training data and feature extraction mode, the case where for not occurring in training data
It is susceptible to flase drop, or the variations such as shines, blocks due to target light and cause missing inspection.And the detection method is not regarded using driving
The prior information that position occurs in the traffic sign of frequency is improved detector.The document is to the traffic sign detected using card
Kalman Filtering method carries out the tracking based on motion model, i.e., using testing result as the observation of Kalman filter, thus
Next frame position of traffic sign is estimated.And this track algorithm is merely with motion model information, once motion model
Suddenly change may result in tracking drift.And the tracking failure caused by variation such as it can not handle the illumination of traffic sign, block
Problem.Identification for traffic sign, existing method do not consider the identification of the different scale size of traffic sign during tracking
As a result to the Weight of the ballot of final result, the robustness of Traffic Sign Recognition cannot be improved.
Invention content
In order to which the detection and recognition methods poor robustness, the present invention that overcome the shortcomings of existing traffic sign provide one kind and are based on
The detection and recognition methods of traffic sign in the driving video of incremental learning.This method is trained using converging channels feature
Adaboost graders, and detector is improved by the spatial prior distribution of the traffic sign for video of driving a vehicle, then will
The testing result of the detector carries out the tracking based on motion model as the observation of Kalman filter, meanwhile, it is tracking
In the process, new increment SVM detectors are trained online, when former Adaboost detectors since the apparent variation of mark causes to examine
It when dendrometry loses, is detected by the online incremental detector, and is inputted testing result as the observation of Kalman filter,
Achieve the effect that while utilizing motion model and apparent model into line trace.During tracking, to fail to be consecutively detected
Target is filtered, and improves the reliability of tracking result.Finally the tracking result of the same physics traffic sign is based on
The Nearest Neighbor with Weighted Voting of the Gauss weight of scale obtains the final output that final recognition result is system.For traffic sign with
The problems such as during track due to the mutation of the apparent state change of target or motion model, the method for the present invention, which uses, is based on increment
The on-line checking device of study, when the state change of target, in the enterprising Mobile state adjustment of detector model, and with this testing result
Former motion model is updated, the final robustness for improving detection and identification.
The technical solution adopted by the present invention to solve the technical problems is：It is handed in a kind of driving video based on incremental learning
The detection and recognition methods of logical mark, its main feature is that using following steps：
Step 1: training dataset is normalized, and sampled images block, channel characteristics pond is calculated, is calculated using Adaboost
Method trains detector model, the detector to cascade to obtain by 4 layers strong detector, this 4 layers strong detector respectively includes 32,128,
512,2048 Weak Classifiers.Detection target by all 4 layers of strong classifiers is candidate target, and the strong of last layer is divided
Class device model is：
Wherein, αtFor the weight for each Weak Classifier that training obtains, htFor Weak Classifier, T is the number of Weak Classifier.
Step 2: carrying out off-line training to online increment SVM detectors using training data, initialization procedure is completed.
Step 3: carrying out the Parzen window density estimations of Gaussian kernel to the positive sample position of training data, it is close to obtain probability
Spending function is：
Wherein, x is vectorial (x, y), i.e., the position coordinates that traffic sign occurs.N is training data positive sample number, VnFor
The window size of window function, hnFor the length of side of window.Function is Gaussian function.I.e.：
Step 4: being handled using following formula the strong classifier model in step 1：
Parameter lambda is determined by the cross-validation experiments on training set.
Step 5: being detected to input video using the detector model that step 2 determines, testing result position is made
Kalman filter is inputted for observation, the tracking to traffic sign is realized by the update of Kalman filtering and forecast period.
Step 6: the tracking result to step 5 carries out positive sample possibility test, test function is as follows：
Lpositive(fk)=symm (fk)·pph(fkmean,fk)·ppos(fk-1,fk) (5)
The high conduct tracking result output of positive sample possibility, and with the parameter of result update Kalman filtering and increasing
Amount update on-line checking device.Wherein Lpositive(fk) be the inverse that kth frame becomes positive sample possibility size, i.e. symmetry is got over
Height, value are smaller.symm(fk) be kth frame target the big small magnitude of symmetry, pph(fkmean,fk) be kth frame target sense
Know the Euclidean distance of Hash and the average value of preceding k frames.ppos(fk-1,fk) be kth frame target and k-1 frame targets coordinate Euclidean
Distance.For the low video frame of positive sample possibility, then using on-line checking device, neighborhood carries out on-line checking in the position, will tie
Fruit is as output as a result, being used to update Kalman filtering parameter simultaneously.
The beneficial effects of the invention are as follows：This method trains Adaboost graders using converging channels feature, and passes through row
The spatial prior distribution of the traffic sign of vehicle video is improved detector, then using the testing result of the detector as card
The observation of Thalmann filter carries out the tracking based on motion model, meanwhile, during tracking, new increment is trained online
SVM detectors, when former Adaboost detectors lead to detection failure due to the apparent variation of mark, by the online incremental detection
Device is detected, and is inputted testing result as the observation of Kalman filter, is reached while being utilized motion model and table
Effect of the sight model into line trace.During tracking, the target for failing to be consecutively detected is filtered, improves tracking result
Reliability.The Nearest Neighbor with Weighted Voting of the Gauss weight based on scale is finally carried out to the tracking result of the same physics traffic sign,
Obtain the final output that final recognition result is system.For the apparent state due to target during traffic sign tracks
The problems such as variation or the mutation of motion model, the method for the present invention uses the on-line checking device based on incremental learning, when the shape of target
When state changes, in the enterprising Mobile state adjustment of detector model, and former motion model is updated with this testing result, is finally carried
The high robustness of detection and identification.
It elaborates with reference to the accompanying drawings and detailed description to the present invention.
Description of the drawings
Fig. 1 is the flow chart of detection and the recognition methods of traffic sign in the driving video the present invention is based on incremental learning.
Fig. 2 is prior probability distribution curve in space in the method for the present invention.
Fig. 3 is the Traffic Sign Recognition accuracy curve that the method for the present invention obtains.
Fig. 4 is that the method for the present invention handles the real shooting photo blocked with illumination variation.
Specific implementation mode
Referring to Fig.1-4.The present invention is based on the detection of traffic sign in the driving video of incremental learning and recognition methods are specific
Steps are as follows：
Step 1, first, training dataset is normalized, and sampled images block, calculates channel characteristics pond, used
Adaboost algorithm trains detector model, the detector to cascade to obtain by 4 layers strong detector, this 4 layers strong detector wraps respectively
Include 32,128,512,2048 Weak Classifiers.Detection target by all 4 layers of strong classifiers is candidate target, and last
One layer of strong classifier model is：
Wherein, αtFor the weight for each Weak Classifier that training obtains, htFor Weak Classifier, T is the number of Weak Classifier.
Step 2, off-line training is carried out to online increment SVM detectors using training data, completes initialization procedure.
Step 3, the Parzen window density estimations that Gaussian kernel is carried out to the positive sample position of training data, obtain probability density
Function is：
Wherein, x is vectorial (x, y), i.e., the position coordinates that traffic sign occurs.N is training data positive sample number, VnFor
The window size of window function, hnFor the length of side of window.Function is Gaussian function.I.e.：
Step 4, the model in step 1 is handled using following formula：
Parameter lambda is determined by the cross-validation experiments on training set.
Step 5, using step 2 determine detector model input video is detected, using testing result position as
Observation inputs Kalman filter, and the tracking to traffic sign is realized by the update of Kalman filtering and forecast period.
Step 6, positive sample possibility test is carried out to the tracking result of step 5, test function is as follows：
Lpositive(fk)=symm (fk)·pph(fkmean,fk)·ppos(fk-1,fk) (10)
The high conduct tracking result output of positive sample possibility, and with the parameter of result update Kalman filtering and increasing
Amount update on-line checking device.Wherein Lpositive(fk) be the inverse that kth frame becomes positive sample possibility size, i.e. symmetry is got over
Height, value are smaller.symm(fk) be kth frame target the big small magnitude of symmetry, pph(fkmean,fk) be kth frame target sense
Know the Euclidean distance of Hash and the average value of preceding k frames.ppos(fk-1,fk) be kth frame target and k-1 frame targets coordinate Euclidean
Distance.For the low video frame of positive sample possibility, then using on-line checking device, neighborhood carries out on-line checking in the position, will tie
Fruit is as output as a result, being used to update Kalman filtering parameter simultaneously.
The effect of the present invention can be described further by following emulation experiment.
1. simulated conditions.
The method of the present invention is to be in central processing unitI5-3470 3.2GHz CPU, memory 4G, WINDOWS 7 are grasped
Make in system, the emulation carried out with MATLAB softwares.
The data used in emulation are public data collection and the data set voluntarily acquired.
2. emulation content.
First, experimental verification spatial prior probabilities are distributed the raising for road traffic sign detection, and experiment knot is shown by Fig. 2
Fruit.
Secondly, obtained Traffic Sign Recognition accuracy Dependence Results are as shown in Figure 3.
Finally, the method that Fig. 4 proves the present invention blocks processing the validity with illumination variation.
As it is clear from fig. 2 that with parameter change, Detection accuracy dramatically increases, and recall rate is substantially unaffected simultaneously, because
This can prove that the distribution of the spatial prior probabilities in the present invention can efficiently reduce flase drop, improve detection performance.
It can be seen that, the Gauss Weight method based on scale in the method for the present invention can be effectively from Fig. 3 curves
Improve recognition accuracy.
The method of the present invention is for traffic sign illumination variation and the validity blocked to a certain degree as seen from Figure 4.
Claims (1)
1. the detection and recognition methods of traffic sign in a kind of driving video based on incremental learning, it is characterised in that including following
Step：
Step 1: training dataset is normalized, and sampled images block, channel characteristics pond is calculated, is instructed using Adaboost algorithm
Practice detector model, which is cascaded to obtain by 4 layers strong detector, this 4 layers strong detector respectively includes 32,128,512,
2048 Weak Classifiers；Detection target by all 4 layers of strong classifiers is candidate target, and the strong classifier of last layer
Model is：
Wherein, αtFor the weight for each Weak Classifier that training obtains, htFor Weak Classifier, T is the number of Weak Classifier；
Step 2: carrying out off-line training to online increment SVM detectors using training data, initialization procedure is completed；
Step 3: carrying out the Parzen window density estimations of Gaussian kernel to the positive sample position of training data, probability density letter is obtained
Number is：
Wherein, x is vectorial (x, y), i.e., the position coordinates that traffic sign occurs；N is training data positive sample number, VnFor window letter
Several window sizes, hnFor the length of side of window；Function is Gaussian function；I.e.：
Step 4: being handled using following formula the strong classifier model in step 1：
Parameter lambda is determined by the cross-validation experiments on training set；
Step 5: being detected to input video using the detector model that step 2 determines, using testing result position as sight
Measured value inputs Kalman filter, and the tracking to traffic sign is realized by the update of Kalman filtering and forecast period；
Step 6: the tracking result to step 5 carries out positive sample possibility test, test function is as follows：
Lpositive(fk)=symm (fk)·pph(fkmean,fk)·ppos(fk-1,fk) (5)
The high conduct tracking result output of positive sample possibility, and more with the parameter of result update Kalman filtering and increment
New on-line checking device；Wherein Lpositive(fk) be the inverse that kth frame becomes positive sample possibility size, i.e. symmetry is higher,
It is worth smaller；symm(fk) be kth frame target the big small magnitude of symmetry, pph(fkmean,fk) be kth frame target perceptual hash
With the Euclidean distance of the average value of preceding k frames；ppos(fk-1,fk) be kth frame target and k-1 frame targets coordinate Euclidean distance；
For the low video frame of positive sample possibility then using on-line checking device in the position neighborhood carry out on-line checking, using result as
Output is as a result, be used to update Kalman filtering parameter simultaneously.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510359501.3A CN104978570B (en) | 2015-06-25 | 2015-06-25 | The detection and recognition methods of traffic sign in driving video based on incremental learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510359501.3A CN104978570B (en) | 2015-06-25 | 2015-06-25 | The detection and recognition methods of traffic sign in driving video based on incremental learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104978570A CN104978570A (en) | 2015-10-14 |
| CN104978570B true CN104978570B (en) | 2018-10-23 |
Family
ID=54275055
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510359501.3A Active CN104978570B (en) | 2015-06-25 | 2015-06-25 | The detection and recognition methods of traffic sign in driving video based on incremental learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104978570B (en) |
Families Citing this family (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017095543A1 (en) * | 2015-12-01 | 2017-06-08 | Intel Corporation | Object detection with adaptive channel features |
| CN107157717A (en) * | 2016-03-07 | 2017-09-15 | 维看公司 | Object detection from visual information to blind person, analysis and prompt system for providing |
| DE102016003424B4 (en) * | 2016-03-21 | 2023-09-28 | Elektrobit Automotive Gmbh | Method and device for recognizing traffic signs |
| CN105787470A (en) * | 2016-03-25 | 2016-07-20 | 黑龙江省电力科学研究院 | Method for detecting power transmission line tower in image based on polymerization multichannel characteristic |
| CN108154102B (en) * | 2017-12-21 | 2021-12-10 | 安徽师范大学 | Road traffic sign identification method |
| CN108805187B (en) * | 2018-05-29 | 2022-07-19 | 北京佳格天地科技有限公司 | Astronomical spectrum sequence automatic classification system and method |
| CN109993207B (en) * | 2019-03-01 | 2022-10-25 | 华南理工大学 | Image privacy protection method and system based on target detection |
| CN110245565A (en) | 2019-05-14 | 2019-09-17 | 东软集团股份有限公司 | Wireless vehicle tracking, device, computer readable storage medium and electronic equipment |
| CN110415270B (en) * | 2019-06-17 | 2020-06-26 | 广东第二师范学院 | Human motion form estimation method based on double-learning mapping incremental dimension reduction model |
| CN110555466A (en) * | 2019-08-13 | 2019-12-10 | 创新奇智(南京)科技有限公司 | Cascade identification network algorithm capable of being dynamically increased |
| CN112686203B (en) * | 2021-01-12 | 2023-10-31 | 重庆大学 | Vehicle safety warning device detection method based on space priori |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101023436A (en) * | 2004-08-16 | 2007-08-22 | 西门子共同研究公司 | Method for traffic sign detection |
| CN102902955A (en) * | 2012-08-30 | 2013-01-30 | 中国科学技术大学 | Method and system for intelligently analyzing vehicle behaviour |
| CN103390167A (en) * | 2013-07-18 | 2013-11-13 | 奇瑞汽车股份有限公司 | Multi-characteristic layered traffic sign identification method |
-
2015
- 2015-06-25 CN CN201510359501.3A patent/CN104978570B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101023436A (en) * | 2004-08-16 | 2007-08-22 | 西门子共同研究公司 | Method for traffic sign detection |
| CN102902955A (en) * | 2012-08-30 | 2013-01-30 | 中国科学技术大学 | Method and system for intelligently analyzing vehicle behaviour |
| CN103390167A (en) * | 2013-07-18 | 2013-11-13 | 奇瑞汽车股份有限公司 | Multi-characteristic layered traffic sign identification method |
Non-Patent Citations (2)
| Title |
|---|
| "一种SVM增量学习算法";萧嵘 等，;《南京大学学报(自然科学)》;20020331;第38卷(第2期);第152-157页 * |
| Andreas Møgelmose etc，."Traffic Sign Detection for U.S. Roads: Remaining Challenges and a case for Tracking".《2014 IEEE 17th International Conference on Intelligent Transportation Systems (ITSC)》.2014,第1394-1399页. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104978570A (en) | 2015-10-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104978570B (en) | The detection and recognition methods of traffic sign in driving video based on incremental learning | |
| Li et al. | Road network extraction via deep learning and line integral convolution | |
| CN104156734B (en) | A kind of complete autonomous on-line study method based on random fern grader | |
| CN103400391A (en) | Multiple-target tracking method and device based on improved random forest | |
| CN108830188A (en) | Vehicle checking method based on deep learning | |
| CN104112282A (en) | A method for tracking a plurality of moving objects in a monitor video based on on-line study | |
| CN107133974A (en) | The vehicle type classification method that Gaussian Background modeling is combined with Recognition with Recurrent Neural Network | |
| CN103793926B (en) | Method for tracking target based on sample reselection procedure | |
| CN105320966A (en) | Vehicle driving state recognition method and apparatus | |
| CN103871077B (en) | A kind of extraction method of key frame in road vehicles monitoring video | |
| CN110070090A (en) | A kind of logistic label information detecting method and system based on handwriting identification | |
| CN102867195B (en) | Method for detecting and identifying a plurality of types of objects in remote sensing image | |
| CN103593672A (en) | Adaboost classifier on-line learning method and Adaboost classifier on-line learning system | |
| CN110348437A (en) | It is a kind of based on Weakly supervised study with block the object detection method of perception | |
| CN105976397B (en) | A kind of method for tracking target | |
| CN102254428A (en) | Traffic jam detection method based on video processing | |
| CN104732248A (en) | Human body target detection method based on Omega shape features | |
| CN102708384B (en) | Bootstrapping weak learning method based on random fern and classifier thereof | |
| CN109376736A (en) | A kind of small video target detection method based on depth convolutional neural networks | |
| Wang et al. | Vehicle reidentification with self-adaptive time windows for real-time travel time estimation | |
| CN104463909A (en) | Visual target tracking method based on credibility combination map model | |
| CN115170611A (en) | Complex intersection vehicle driving track analysis method, system and application | |
| Wu et al. | Single shot multibox detector for vehicles and pedestrians detection and classification | |
| CN103971362B (en) | SAR image change-detection based on rectangular histogram and elite genetic algorithm for clustering | |
| CN104392466A (en) | Detecting and tracking method based on visual salient original target |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |