CN106886757B - A kind of multiclass traffic lights detection method and system based on prior probability image - Google Patents
A kind of multiclass traffic lights detection method and system based on prior probability image Download PDFInfo
- Publication number
- CN106886757B CN106886757B CN201710043802.4A CN201710043802A CN106886757B CN 106886757 B CN106886757 B CN 106886757B CN 201710043802 A CN201710043802 A CN 201710043802A CN 106886757 B CN106886757 B CN 106886757B
- Authority
- CN
- China
- Prior art keywords
- traffic lights
- image
- feature
- sliding window
- positive
- 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
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
- G06V20/584—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of vehicle lights or traffic lights
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Traffic Control Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of multiclass traffic lights detection method and system based on prior probability image, belong to technical field of computer vision.The method of the present invention extracts the integrating channel feature of the positive and negative sample image of traffic lights first, using Adaboost learning algorithm one traffic lights detection model of training, the regularity of distribution of traffic lights is recycled to construct prior probability image, detection threshold value is adaptively arranged, traffic lights positioning is carried out to target image by the way of soft cascade, Classification and Identification finally is carried out to traffic lights again, category label, one multi classifier of training first are carried out to the traffic lights sample of different shape and state before identification；Type identification is carried out to traffic lights image-region again.The invention also discloses a kind of the multiclass traffic lights detection system based on prior probability image, training of the technical solution of the present invention Jing Guo great amount of samples, it is not easy to be influenced by noise and illumination variation, suitable for complicated urban transportation scene.
Description
Technical field
The invention belongs to technical field of computer vision, hand over more particularly, to a kind of multiclass based on prior probability image
Logical lamp inspection surveys method and system.
Background technique
With increasing for automobile, traffic problems are increasingly prominent.In order to solve traffic congestion, reduction accident rate is more next
More research institutions have put into the research of intelligent transportation system, and the hub device that traffic lights is operated as traffic, are
Wherein essential a part.If energy is timely, is automatically captured traffic light signals, either to unmanned or intelligence
It can assist driving, there is very important meaning.
Traffic lights detection system generallys use short range communication, global positioning system (Global Positioning
System, GPS), the modes such as computer vision.Wherein, using the detection method of computer vision since it is easily installed, cost
Cheap, effect is intuitive, and is taken seriously without the features such as increasing additional infrastructure.This method on automobile by installing one
A video camera acquires the image of vehicle front road scene, analyzes and whether there is traffic lights in image, and to its position and type
Judged.
Currently based on the traffic lights detection method of computer vision substantially using didactic method, it is generally divided into
Following steps: obtaining the candidate region of traffic lights first, is usually split in specific color space to image,
Or the speck region in image is obtained using top cap algorithm；Secondly template matching or connected domain are adopted to obtained candidate region
The modes such as analysis are filtered, and obtain final traffic lights testing result.
Although above method speed ratio is very fast, since its detection mode is the rule using some artificial settings, answer
Range is usually relatively more limited.Traffic light kind is numerous, and being difficult design one can be accurate to all types of traffic lights
The method of detection, so existing most of method can be only applied to circular traffic lights.And the above method is to illumination variation
And the poor robustness of noise, due to the blocking of traffic lights, other lamps in illumination variation, halation, flashing and complex scene
The presence of the problems such as interference, the model being calculated often may be only available for some specific data set, can not be applied to complexity
Real scene in.
Summary of the invention
Aiming at the above defects or improvement requirements of the prior art, the present invention provides a kind of multiclass based on prior probability image
Traffic lights detection method and system, its object is to use the traffic lights detection method based on study, to different in collecting sample
The traffic lights sample of shape is marked, and using multi classifier one multi classifier of training, recycles the distribution of traffic lights
Rule building prior probability image, is adaptively arranged classification thresholds, and using integrating channel feature as traffic lights mould
Thus the feature representation of type solves often to may be only available for some specific shape traffic lights and anti-interference energy in prior art
The technical issues of power.
To achieve the above object, according to one aspect of the present invention, a kind of multiclass friendship based on prior probability image is provided
Logical lamp detection method, method includes the following steps:
(1) according to the ratio set to testing image into multiple dimensioned scaling；Preferably, Image scaling coefficients first layer is
1, each layer is the 0.95 of upper one layer later, scales 20 layers in total；
(2) to all image zooming-out integrating channel features after scaling, feature pyramid is formed；
(3) it is slided using the window of a fixed size in the pyramidal each layer of feature, and utilizes traffic lamp inspection
It surveys model to detect the integrating channel feature in sliding window, obtains the sliding window position and window comprising traffic lights
Traffic lights probability score；
(4) statistics is comprising traffic lights and the overlapping area of overlapped sliding window, overlapping area are more than the weight of setting
Folded threshold value then chooses the highest sliding window of traffic lights probability score as window to be sorted；Preferably, anti-eclipse threshold 0.5；
(5) window to be sorted is input to traffic lights disaggregated model, obtains the type of traffic lights and state and exported.
Further, in the step (3) traffic lights detection model training specifically includes the following steps:
(21) road scene image is acquired, to include all traffic light kinds that need to be detected in collected sample image
And state, and different type and the number of state traffic lights are evenly distributed；
(22) handmarking being carried out to image pattern, region of the cutting comprising complete traffic lights is simultaneously labeled as positive sample, with
Machine, which is chosen, does not include traffic lights region as negative sample, and positive and negative sample-size is normalized；
(23) the feature channel for choosing positive negative sample, extracts the characteristics of image in positive and negative sample characteristics channel later, finally right
Obtained characteristics of image is gathered the integrating channel feature for obtaining positive negative sample；
(24) the integrating channel feature of the positive negative sample of traffic lights is utilized, one Adaboost classifier of training is as traffic lights
Detection model.
Further, the integrating channel feature in sliding window is detected specifically using soft grade in the step (3)
Join Adaboost, Adaboost classifier is made of multiple Weak Classifiers, such as following formula
Wherein, x is to be extracted integrating channel feature vector in sliding window；hi(x) for i-th Weak Classifier to feature to
Measure the output of x；Since first Weak Classifier, compare first Weak Classifier and self-adapting detecting threshold value, if being less than adaptive
Detection threshold value is answered, then assert it is background in sliding window, continues sliding window and is detected next time；If more than self-adapting detecting
Threshold value then adds up the latter Weak Classifier, and accumulation result is continued and self-adapting detecting threshold comparison, if being less than adaptive inspection
Threshold value is surveyed, then assert that sliding window is background, continues sliding window and is detected next time；Then if more than self-adapting detecting threshold value
Continue cumulative the latter Weak Classifier, until by the last one Weak Classifier, then assert that the sliding window includes traffic lights mesh
Mark exports accumulation result as traffic lights probability score；Export sliding window position.
Further, the self-adapting detecting threshold value seek the following steps are included:
(41) the traffic lights location-prior probability graph and traffic lights size prior probability image of all positive samples are calculated；
(42) calculating position prior probability image IpIntegrogram
Wherein, the value of each point of integrogram is equivalent in original image in the cumulative of this upper left corner all values；By right
Integrogram carries out simple arithmetical operation, obtains
Wherein, x and y indicates that the coordinate in the detection window upper left corner, w and h respectively indicate the width and height of detection window；
(43) by size prior probability image Is, calculate candidate area size and the region occur traffic lights desired size it
Between deviation Ps(x,y,w,h)
(44) self-adapting detecting threshold value is
T (x, y, w, h)=- Ps(x,y,w,h)·Pp(x,y,w,h)。
Further, the training of traffic lights disaggregated model is specifically divided into following steps in the step (5):
(51) manual sort's label is carried out to the traffic lights of different shape different conditions in positive sample；
(52) the feature channel for choosing positive sample, extracts the characteristics of image in positive sample feature channel later, finally to obtaining
Characteristics of image gathered the integrating channel feature for obtaining positive negative sample；
(53) it using classification positive sample, is trained using multi classifier, obtains traffic lights disaggregated model；Preferably,
Multi classifier uses Adaboost.MH algorithm.
It is another aspect of this invention to provide that a kind of multiclass traffic lights detection system based on prior probability image is provided, it should
System includes following part:
Testing image Zoom module, for according to the ratio set to testing image into multiple dimensioned scaling；Preferably, scheme
Picture zoom factor first layer is 1, and each layer is the 0.95 of upper one layer later, scales 20 layers in total；
Integrating channel characteristic extracting module, for forming feature to all image zooming-out integrating channel features after scaling
Pyramid；
Detection module is slided, for being slided using the window of a fixed size in the pyramidal each layer of feature,
And the integrating channel feature in sliding window is detected using traffic lights detection model, obtain the sliding window comprising traffic lights
The traffic lights probability score of mouth position and window；
Non-maxima suppression module includes traffic lights and the overlapping area of overlapped sliding window for counting, weight
Folded area is more than that the anti-eclipse threshold of setting then chooses the highest sliding window of traffic lights probability score as window to be sorted；It is preferred that
, anti-eclipse threshold 0.5；
Categorization module obtains the type and state of traffic lights for window to be sorted to be input to traffic lights disaggregated model
And it exports.
Further, the training system of traffic lights detection model is specifically included with lower unit in the sliding detection module:
Sampling unit will own for acquiring road scene image in collected sample image comprising what need to be detected
Traffic light kind and state, and different type and the number of state traffic lights are evenly distributed；
Positive and negative sample labeling unit, for carrying out handmarking to image pattern, cutting includes the region of complete traffic lights
And it is labeled as positive sample, it randomly selects not comprising traffic lights region as negative sample, positive and negative sample-size is normalized；
Integrating channel feature extraction unit extracts positive and negative sample characteristics for choosing the feature channel of positive negative sample later
The characteristics of image in channel is finally gathered the integrating channel feature for obtaining positive negative sample to obtained characteristics of image；
Traffic lights detection model training unit, for the integrating channel feature using the positive negative sample of traffic lights, training one
Adaboost classifier is as traffic lights detection model.
Further, the integrating channel feature in sliding window is carried out detecting specific use in the sliding detection module
Soft cascade Adaboost, Adaboost classifier is made of multiple Weak Classifiers, such as following formula
Wherein, x is to be extracted integrating channel feature vector in sliding window；hi(x) for i-th Weak Classifier to feature to
Measure the output of x；Since first Weak Classifier, compare first Weak Classifier and self-adapting detecting threshold value, if being less than adaptive
Detection threshold value is answered, then assert it is background in sliding window, continues sliding window and is detected next time；If more than self-adapting detecting
Threshold value then adds up the latter Weak Classifier, and accumulation result is continued and self-adapting detecting threshold comparison, if being less than adaptive inspection
Threshold value is surveyed, then assert that sliding window is background, continues sliding window and is detected next time；Then if more than self-adapting detecting threshold value
Continue cumulative the latter Weak Classifier, until by the last one Weak Classifier, then assert that the sliding window includes traffic lights mesh
Mark exports accumulation result as traffic lights probability score；Export sliding window position.
Further, the system of seeking of the self-adapting detecting threshold value includes with lower unit:
Prior probability image computing unit, for calculating the traffic lights location-prior probability graph and traffic lights ruler of all positive samples
Very little prior probability image；
Integrogram computing unit is used for calculating position prior probability image IpIntegrogram
Wherein, the value of each point of integrogram is equivalent in original image in the cumulative of this upper left corner all values；By right
Integrogram carries out simple arithmetical operation, obtains
Wherein, x and y indicates that the coordinate in the detection window upper left corner, w and h respectively indicate the width and height of detection window；
Deviation computing unit, for by size prior probability image Is, calculate candidate area size and traffic occur in the region
Deviation P between the desired size of lamps(x,y,w,h)
Self-adapting detecting threshold computation unit, for calculating self-adapting detecting threshold value
T (x, y, w, h)=- Ps(x,y,w,h)·Pp(x,y,w,h)。
Further, the training system of traffic lights disaggregated model is specifically divided into lower unit in the categorization module:
Classification marker unit carries out manual sort's label for the traffic lights to different shape different conditions in positive sample；
Integrating channel feature extraction unit of classifying extracts positive sample feature for choosing the feature channel of positive sample later
The characteristics of image in channel is finally gathered the integrating channel feature for obtaining positive negative sample to obtained characteristics of image；
Traffic lights disaggregated model training unit, for being trained, being obtained using multi classifier using classification positive sample
Traffic lights disaggregated model, it is preferred that multi classifier uses Adaboost.MH algorithm.
In general, through the invention it is contemplated above technical scheme is compared with the prior art, have following technology special
Sign and the utility model has the advantages that
(1) the traffic lights detection method based on study is used, traffic lights sample of different shapes in collecting sample is carried out
Label realizes the identification to different type traffic lights, it can be determined that inspection using multi classifier one multi classifier of training
The shape and state of the traffic lights measured；
(2) present invention uses feature representation of the integrating channel feature as traffic lights model.It is special with simple use color
Sign or Gradient Features carry out target detection comparatively, integrating channel feature is extracted a variety of different types of features again first
It is merged, so as to form the polynary expression of target, enhances the robustness of feature；
(3) by the way of adaptive threshold, in conjunction with the characteristic distributions of traffic lights, traffic lights probability of occurrence and size are utilized
And the threshold value of the relationship setting detection between picture position, improve the precision of detection.
Detailed description of the invention
Fig. 1 overall flow figure of the present invention；
Image channel schematic diagram in Fig. 2 present invention；
Location-prior probability schematic diagram in Fig. 3 a present invention；
Size prior probability schematic diagram in Fig. 3 b present invention；
Traffic lights positioning function schematic diagram in Fig. 4 a, 4b, 4c and 4d present invention.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as in addition, technical characteristic involved in the various embodiments of the present invention described below
Not constituting a conflict with each other can be combined with each other.
It is as shown in Figure 1 the implementing procedure of the embodiment of the present invention, specifically includes the following steps:
(1) according to the ratio set to testing image into multiple dimensioned scaling；Zoom factor first layer is 1 in this implementation,
Each layer is the 0.95 of upper one layer later, scales 20 layers in total；
(2) to all image zooming-out integrating channel features after scaling, feature pyramid is formed；Image after scaling is mentioned
Integrating channel feature is taken, feature pyramid is formed；
Channel is the output response of original image, indicates extracted certain type of feature；For color image,
Each Color Channel then corresponds to a channel, and other similar channel can carry out various linear and nonlinear transformation to original image
It obtains；The color and gradient feature of this Cass collection traffic lights constructs ten feature channels, as shown in Fig. 2, wherein including
Tri- Color Channels of LUV, a gradient intensity channel and gradient projection are at 0 degree to six direction equally distributed between 180 degree
The six gradient direction channels constituted；On feature channel extract Haar feature, by the Haar feature in all feature channels to
Amount connects to obtain the integrating channel feature of sample.
(3) it is slided using the window of a fixed size in the pyramidal each layer of feature, and utilizes traffic lamp inspection
It surveys model to detect the integrating channel feature in sliding window, obtains the sliding window position and window comprising traffic lights
Traffic lights probability score；
Wherein traffic lights detection model training specifically includes the following steps:
(21) road scene image is acquired, to include all traffic light kinds that need to be detected in collected sample image
And state, and different type and the number of state traffic lights are evenly distributed；
(22) handmarking being carried out to image pattern, region of the cutting comprising complete traffic lights is simultaneously labeled as positive sample, with
Machine, which is chosen, does not include traffic lights region as negative sample, and positive and negative sample-size is normalized；
(23) the feature channel for choosing positive negative sample, extracts the characteristics of image in positive and negative sample characteristics channel later, finally right
Obtained characteristics of image is gathered the integrating channel feature for obtaining positive negative sample；
(24) the integrating channel feature of the positive negative sample of traffic lights is utilized, one Adaboost classifier of training is as traffic lights
Detection model.
In order to accelerate the speed of detection, present invention employs soft cascade Adaboost, Adaboost classifier is by multiple
Weak Classifier composition
Wherein, x is to be extracted integrating channel feature vector in sliding window；hi(x) for i-th Weak Classifier to feature to
Measure the output of x；Since first Weak Classifier, compare first Weak Classifier and self-adapting detecting threshold value, if being less than adaptive
Detection threshold value is answered, then assert it is background in sliding window, continues sliding window and is detected next time；If more than self-adapting detecting
Threshold value then adds up the latter Weak Classifier, and accumulation result is continued and self-adapting detecting threshold comparison, if being less than adaptive inspection
Threshold value is surveyed, then assert that sliding window is background, continues sliding window and is detected next time；Then if more than self-adapting detecting threshold value
Continue cumulative the latter Weak Classifier, until by the last one Weak Classifier, then assert that the sliding window includes traffic lights mesh
Mark exports accumulation result as traffic lights probability score；Export sliding window position.
Wherein self-adapting detecting threshold value seek the following steps are included:
(41) the traffic lights location-prior probability graph and traffic lights size prior probability image of all positive samples are calculated；
(42) calculating position prior probability image IpIntegrogram
Wherein, the value of each point of integrogram is equivalent in original image in the cumulative of this upper left corner all values；By right
Integrogram carries out simple arithmetical operation, obtains
Wherein, x and y indicates that the coordinate in the detection window upper left corner, w and h respectively indicate the width and height of detection window；
(43) by size prior probability image Is, calculate candidate area size and the region occur traffic lights desired size it
Between deviation Ps(x,y,w,h)
(44) self-adapting detecting threshold value is
T (x, y, w, h)=- Ps(x,y,w,h)·Pp(x,y,w,h)。
(4) statistics is comprising traffic lights and the overlapping area of overlapped sliding window, overlapping area are more than the weight of setting
Folded threshold value then chooses the highest sliding window of traffic lights probability score as window to be sorted；The present embodiment anti-eclipse threshold is 0.5.
(5) window to be sorted is input to traffic lights disaggregated model, obtains the type of traffic lights and state and exported；
Wherein traffic lights disaggregated model training the following steps are included:
(51) manual sort's label is carried out to the traffic lights of different shape different conditions in positive sample；
(52) the feature channel for choosing positive sample, extracts the characteristics of image in positive sample feature channel later, finally to obtaining
Characteristics of image gathered the integrating channel feature for obtaining positive negative sample；
(53) it using classification positive sample, is trained using multi classifier, obtains traffic lights disaggregated model；The present embodiment
Multi classifier uses Adaboost.MH algorithm.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (6)
1. a kind of multiclass traffic lights detection method based on prior probability image, which is characterized in that method includes the following steps:
(1) according to the ratio set to testing image into multiple dimensioned scaling；
(2) to all image zooming-out integrating channel features after scaling, feature pyramid is formed；
(3) it is slided using the window of a fixed size in the pyramidal each layer of feature, and detects mould using traffic lights
Type detects the integrating channel feature in sliding window, obtains the traffic of the sliding window position comprising traffic lights and window
Lamp probability score；
The integrating channel feature in sliding window detect in the step (3) and specifically uses soft cascade Adaboost,
Adaboost classifier is made of multiple Weak Classifiers, such as following formula
Wherein, x is to be extracted integrating channel feature vector in sliding window；hiIt (x) is i-th of Weak Classifier to feature vector x's
Output；Since first Weak Classifier, compare first Weak Classifier and self-adapting detecting threshold value, if being less than self-adapting detecting
Threshold value then assert it is background in sliding window, continues sliding window and is detected next time；Then if more than self-adapting detecting threshold value
Cumulative the latter Weak Classifier, and by accumulation result continuation and self-adapting detecting threshold comparison, if being less than self-adapting detecting threshold value,
Then assert that sliding window is background, continues sliding window and detected next time；Then continue to tire out if more than self-adapting detecting threshold value
Add the latter Weak Classifier, until by the last one Weak Classifier, then assert that the sliding window includes traffic light target, output
Accumulation result is as traffic lights probability score；Export sliding window position；Wherein, the self-adapting detecting threshold value is sought wrapping
Include following steps:
(41) the traffic lights location-prior probability graph and traffic lights size prior probability image of all positive samples are calculated；
(42) calculating position prior probability image IpIntegrogram
Wherein, the value of each point of integrogram is equivalent in original image in the cumulative of this upper left corner all values；By to integral
Figure carries out simple arithmetical operation, obtains
Wherein, x and y indicates that the coordinate in the detection window upper left corner, w and h respectively indicate the width and height of detection window；
(43) by size prior probability image Is, calculate candidate area size and the region occur between the desired size of traffic lights
Deviation Ps(x,y,w,h)
(44) self-adapting detecting threshold value is
T (x, y, w, h)=- Ps(x,y,w,h)·Pp(x,y,w,h)；
(4) statistics is comprising traffic lights and the overlapping area of overlapped sliding window, overlapping area are more than the overlapping threshold of setting
Value then chooses the highest sliding window of traffic lights probability score as window to be sorted；
(5) window to be sorted is input to traffic lights disaggregated model, obtains the type of traffic lights and state and exported.
2. a kind of multiclass traffic lights detection method based on prior probability image according to claim 1, which is characterized in that institute
State the training of traffic lights detection model in step (3) specifically includes the following steps:
(21) road scene image is acquired, to include all traffic light kinds and shape that need to be detected in collected sample image
State, and different type and the number of state traffic lights are evenly distributed；
(22) handmarking is carried out to image pattern, region of the cutting comprising complete traffic lights is simultaneously labeled as positive sample, random to select
It takes and does not include traffic lights region as negative sample, positive and negative sample-size is normalized；
(23) the feature channel for choosing positive negative sample, extracts the characteristics of image in positive and negative sample characteristics channel later, finally to obtaining
Characteristics of image gathered the integrating channel feature for obtaining positive negative sample；
(24) the integrating channel feature of the positive negative sample of traffic lights is utilized, one Adaboost classifier of training is detected as traffic lights
Model.
3. a kind of multiclass traffic lights detection method based on prior probability image according to claim 1 or 2, feature exist
In the training of traffic lights disaggregated model is specifically divided into following steps in the step (5):
(51) manual sort's label is carried out to the traffic lights of different shape different conditions in positive sample；
(52) the feature channel for choosing positive sample, extracts the characteristics of image in positive sample feature channel, finally to obtained figure later
As feature is gathered the integrating channel feature for obtaining positive negative sample；
(53) it using classification positive sample, is trained using multi classifier, obtains traffic lights disaggregated model.
4. a kind of multiclass traffic lights detection system based on prior probability image, which is characterized in that the system comprises the following modules:
Testing image Zoom module, for according to the ratio set to testing image into multiple dimensioned scaling；
Integrating channel characteristic extracting module, for forming feature gold word to all image zooming-out integrating channel features after scaling
Tower；
Detection module is slided, for being slided using the window of a fixed size in the pyramidal each layer of feature, and benefit
The integrating channel feature in sliding window is detected with traffic lights detection model, obtains the sliding window position comprising traffic lights
Set the traffic lights probability score with window；
The integrating channel feature in sliding window detect specifically using soft cascade in the sliding detection module
Adaboost, Adaboost classifier are made of multiple Weak Classifiers, such as following formula
Wherein, x is to be extracted integrating channel feature vector in sliding window；hiIt (x) is i-th of Weak Classifier to feature vector x's
Output；Since first Weak Classifier, compare first Weak Classifier and self-adapting detecting threshold value, if being less than self-adapting detecting
Threshold value then assert it is background in sliding window, continues sliding window and is detected next time；Then if more than self-adapting detecting threshold value
Cumulative the latter Weak Classifier, and by accumulation result continuation and self-adapting detecting threshold comparison, if being less than self-adapting detecting threshold value,
Then assert that sliding window is background, continues sliding window and detected next time；Then continue to tire out if more than self-adapting detecting threshold value
Add the latter Weak Classifier, until by the last one Weak Classifier, then assert that the sliding window includes traffic light target, output
Accumulation result is as traffic lights probability score；Export sliding window position；
The system of seeking of the self-adapting detecting threshold value includes with lower unit:
Prior probability image computing unit, the traffic lights location-prior probability graph and traffic lights size for calculating all positive samples are first
Test probability graph；
Integrogram computing unit is used for calculating position prior probability image IpIntegrogram
Wherein, the value of each point of integrogram is equivalent in original image in the cumulative of this upper left corner all values；By to integral
Figure carries out simple arithmetical operation, obtains
Wherein, x and y indicates that the coordinate in the detection window upper left corner, w and h respectively indicate the width and height of detection window；
Deviation computing unit, for by size prior probability image Is, calculate candidate area size and the phase of traffic lights occur in the region
Hope the deviation P between sizes(x,y,w,h)
Self-adapting detecting threshold computation unit, for calculating self-adapting detecting threshold value
T (x, y, w, h)=- Ps(x,y,w,h)·Pp(x,y,w,h)；
Non-maxima suppression module includes traffic lights and the overlapping area of overlapped sliding window for counting, faying surface
Product is more than that the anti-eclipse threshold of setting then chooses the highest sliding window of traffic lights probability score as window to be sorted；
Categorization module obtains the type and state and defeated of traffic lights for window to be sorted to be input to traffic lights disaggregated model
Out.
5. a kind of multiclass traffic lights detection system based on prior probability image according to claim 4, which is characterized in that institute
The training system for stating traffic lights detection model in sliding detection module is specifically included with lower unit:
Sampling unit will include all traffic that need to be detected in collected sample image for acquiring road scene image
Lamp type and state, and different type and the number of state traffic lights are evenly distributed；
Positive and negative sample labeling unit, for carrying out handmarking, region of the cutting comprising complete traffic lights and mark to image pattern
It is denoted as positive sample, randomly selects not comprising traffic lights region as negative sample, positive and negative sample-size is normalized；
Integrating channel feature extraction unit extracts positive and negative sample characteristics channel for choosing the feature channel of positive negative sample later
Characteristics of image, the integrating channel feature for obtaining positive negative sample finally is gathered to obtained characteristics of image；
Traffic lights detection model training unit, for the integrating channel feature using the positive negative sample of traffic lights, training one
Adaboost classifier is as traffic lights detection model.
6. a kind of multiclass traffic lights detection system based on prior probability image according to claim 4 or 5, feature exist
In the training system of traffic lights disaggregated model is specifically divided into lower unit in the categorization module:
Classification marker unit carries out manual sort's label for the traffic lights to different shape different conditions in positive sample；
Integrating channel feature extraction unit of classifying extracts positive sample feature channel for choosing the feature channel of positive sample later
Characteristics of image, the integrating channel feature for obtaining positive negative sample finally is gathered to obtained characteristics of image；
Traffic lights disaggregated model training unit, for being trained using multi classifier, obtaining traffic using classification positive sample
Lamp disaggregated model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710043802.4A CN106886757B (en) | 2017-01-19 | 2017-01-19 | A kind of multiclass traffic lights detection method and system based on prior probability image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710043802.4A CN106886757B (en) | 2017-01-19 | 2017-01-19 | A kind of multiclass traffic lights detection method and system based on prior probability image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106886757A CN106886757A (en) | 2017-06-23 |
| CN106886757B true CN106886757B (en) | 2019-06-11 |
Family
ID=59176706
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710043802.4A Active CN106886757B (en) | 2017-01-19 | 2017-01-19 | A kind of multiclass traffic lights detection method and system based on prior probability image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106886757B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108009554A (en) * | 2017-12-01 | 2018-05-08 | 国信优易数据有限公司 | A kind of image processing method and device |
| CN109646796A (en) * | 2019-01-17 | 2019-04-19 | 浙江大学 | Channel wireless radio multi closed loop stimulation system for epilepsy therapy |
| CN110211650A (en) * | 2019-05-30 | 2019-09-06 | 苏州爱医斯坦智能科技有限公司 | Operation Rich Media's electronic health record monitors the method and device of identification record automatically |
| US20210211568A1 (en) * | 2020-01-07 | 2021-07-08 | Motional Ad Llc | Systems and methods for traffic light detection |
| CN112633313B (en) * | 2020-10-13 | 2021-12-03 | 北京匠数科技有限公司 | Bad information identification method of network terminal and local area network terminal equipment |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103886308A (en) * | 2014-04-15 | 2014-06-25 | 中南大学 | Pedestrian detection method through soft cascade classifiers according to polymerization channel characteristics |
| CN105046197A (en) * | 2015-06-11 | 2015-11-11 | 西安电子科技大学 | Multi-template pedestrian detection method based on cluster |
| CN105184278A (en) * | 2015-09-30 | 2015-12-23 | 深圳市商汤科技有限公司 | Human face detection method and device |
| CN105550675A (en) * | 2016-02-02 | 2016-05-04 | 天津大学 | Binocular pedestrian detection method based on optimization polymerization integration channel |
| 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 |
-
2017
- 2017-01-19 CN CN201710043802.4A patent/CN106886757B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103886308A (en) * | 2014-04-15 | 2014-06-25 | 中南大学 | Pedestrian detection method through soft cascade classifiers according to polymerization channel characteristics |
| CN105046197A (en) * | 2015-06-11 | 2015-11-11 | 西安电子科技大学 | Multi-template pedestrian detection method based on cluster |
| CN105184278A (en) * | 2015-09-30 | 2015-12-23 | 深圳市商汤科技有限公司 | Human face detection method and device |
| CN105550675A (en) * | 2016-02-02 | 2016-05-04 | 天津大学 | Binocular pedestrian detection method based on optimization polymerization integration channel |
| 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 (1)
| Title |
|---|
| "基于车载摄像头的行人检测与跟踪方法研究";姚栖;《中国优秀硕士学位论文全文数据库 信息科技辑》;20150715(第7期);正文第2.1.2节，第4.1节，第4.3.1节 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106886757A (en) | 2017-06-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106886757B (en) | A kind of multiclass traffic lights detection method and system based on prior probability image | |
| CN108830188B (en) | Vehicle detection method based on deep learning | |
| CN105160309B (en) | Three lanes detection method based on morphological image segmentation and region growing | |
| CN109902806A (en) | Method is determined based on the noise image object boundary frame of convolutional neural networks | |
| CN104050827B (en) | A kind of traffic lights of view-based access control model detect recognition methods automatically | |
| CN103034836B (en) | Road sign detection method and road sign checkout equipment | |
| CN103902976B (en) | A kind of pedestrian detection method based on infrared image | |
| Fang et al. | Road-sign detection and tracking | |
| CN109190444B (en) | Method for realizing video-based toll lane vehicle feature recognition system | |
| CN111553201B (en) | Traffic light detection method based on YOLOv3 optimization algorithm | |
| CN108921083B (en) | Illegal mobile vendor identification method based on deep learning target detection | |
| CN108171136A (en) | A kind of multitask bayonet vehicle is to scheme to search the system and method for figure | |
| CN103824081B (en) | Method for detecting rapid robustness traffic signs on outdoor bad illumination condition | |
| CN105809121A (en) | Multi-characteristic synergic traffic sign detection and identification method | |
| CN103810503A (en) | Depth study based method for detecting salient regions in natural image | |
| CN102968646A (en) | Plate number detecting method based on machine learning | |
| Wang et al. | License plate detection using gradient information and cascade detectors | |
| CN106650731A (en) | Robust license plate and logo recognition method | |
| CN104268528A (en) | Method and device for detecting crowd gathered region | |
| CN103473571A (en) | Human detection method | |
| CN107704853A (en) | A kind of recognition methods of the traffic lights based on multi-categorizer | |
| CN104021375A (en) | Model identification method based on machine learning | |
| CN104657724A (en) | Method for detecting pedestrians in traffic videos | |
| CN111914839A (en) | Synchronous end-to-end license plate positioning and identifying method based on YOLOv3 | |
| CN110032935A (en) | A kind of traffic signals label detection recognition methods based on deep learning cascade network |
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 |