CN107016390B - A kind of vehicle part detection method and system based on relative position - Google Patents
A kind of vehicle part detection method and system based on relative position Download PDFInfo
- Publication number
- CN107016390B CN107016390B CN201710234936.4A CN201710234936A CN107016390B CN 107016390 B CN107016390 B CN 107016390B CN 201710234936 A CN201710234936 A CN 201710234936A CN 107016390 B CN107016390 B CN 107016390B
- Authority
- CN
- China
- Prior art keywords
- interest
- area
- vehicle part
- relative position
- vehicle
- 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
- G06V20/60—Type of objects
- G06V20/62—Text, e.g. of license plates, overlay texts or captions on TV images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
- G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
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)
- Image Analysis (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention discloses a kind of vehicle part detection method and system based on relative position, wherein the realization of method includes off-line training and on-line checking part, off-line training part includes: acquisition vehicle sample image, extracts the relative position of the sample area-of-interest of sample vehicle part；Gauss model is established using relative position, obtains prior information, using prior information more raw score of the new samples area-of-interest in fast convolution neural network, thus obtains trained relative position network；On-line checking part includes: input vehicle image, extracts the area-of-interest of vehicle part, area-of-interest input relative position network is obtained the score of area-of-interest, the area-of-interest of highest scoring is vehicle part target area.Vehicle part detection is carried out the present invention is based on the relative position information of vehicle part and significantly improves the reliability of vehicle part detection, further enhances the accuracy of vehicle part detection.
Description
Technical field
The invention belongs to area of pattern recognition, more particularly, to a kind of vehicle part detection side based on relative position
Method and system.
Background technique
In each field of intelligent transportation system, often it is related to the detection of road vehicle.According to the difference of utilized data,
Road vehicle test problems can be divided into based on sequence of video images and based on two class basic problem of still image.For the former,
Since there is a large amount of relevant informations to be used between frame each in sequence of video images, so upper opposite in processing
It is easier to realize.And the realization of the vehicle detecting algorithm based on still image is then much more difficult, but due to its higher science
Value and more extensive application prospect, have attracted more and more attention from people the research of this problem, also become intelligent transportation
An important branch in system research.For example, in crossing traffic condition monitoring, parking lot dispatching and monitoring, accident detection, automatic
Still image vehicle detection suffers from broad application prospect in navigation etc. system.
For vehicle component detection, traditional detection method mainly with characteristics of image (such as HOG feature, LBP feature and
SIFT feature etc.) extract based on, classification is then using support vector machines and Adaboost algorithm etc., these methods are due to mentioning
The feature taken is more unilateral, without the characteristic feature of of overall importance and can not be adaptive extraction image, therefore accuracy rate is generally not
It is too high.
Summary of the invention
Aiming at the above defects or improvement requirements of the prior art, the present invention provides a kind of vehicle portions based on relative position
Part detection method and system, its object is to training relative position network when the relative position based on vehicle part obtain priori
Information updates score using prior information, the precision of relative position network is effectively increased, to improve the target detected
The accuracy rate in region.
To achieve the above object, according to one aspect of the present invention, a kind of vehicle part based on relative position is provided
Detection method, including off-line training and on-line checking part,
Off-line training part includes:
(1) vehicle sample image is acquired, the area-of-interest of sample vehicle part is extracted, selectes a sample vehicle part
Area-of-interest on the basis of, record the relative position between the area-of-interest and benchmark of other sample vehicle parts；
(2) Gauss model is established using relative position, obtains prior information, utilize prior information more new samples region of interest
Thus raw score of the domain in fast convolution neural network obtains trained relative position network；
On-line checking part includes:
Vehicle image is inputted, the area-of-interest of vehicle part is extracted, area-of-interest input relative position network is obtained
To the score of area-of-interest, the area-of-interest of highest scoring is vehicle part target area.
Further, the specific implementation of step (2) are as follows:
The relative position (2-1) includes: relative angle and relative distance, establishes Gauss model using relative position, obtains elder generation
Test information:
Wherein, Δ1(x) prior information of the relative angle of the area-of-interest of vehicle part, Δ are indicated2(x) vehicle is indicated
The prior information of the relative distance of the area-of-interest of component, wherein Aj、Bj、CjParameter for the Gaussian function for needing to be fitted, L1
(x) relative angle of the area-of-interest of vehicle part, L are indicated2(x) relative angle of the area-of-interest of vehicle part is indicated
Degree, x indicate area-of-interest；
(2-2) updates score of the area-of-interest in fast convolution neural network using prior information:
P*=arg max (Δ (x) s (x))
Pi *Indicate the score of the area-of-interest of vehicle part, Δ (x) indicates the opposite of the area-of-interest of vehicle part
The prior information of position, s (x) indicate that raw score of the area-of-interest of vehicle part in fast convolution neural network, α are
The weight of relative angle prior information, β are the weight of the prior information of relative distance.
Further, the fitting of the parameter of Gaussian function includes:
As j=1,
Equipped with one group of experimental data (xi, yi) (i=1,2,3 ... n), xiIt is one of the area-of-interest of vehicle part
Relative angle, yiFor the prior information of the relative position of the area-of-interest of vehicle part, then formula (I) can be written as:
Above formula both sides take natural logrithm, turn to
It enables
Consider total Test data, then indicate in the matrix form are as follows:
It is abbreviated as
Z=XB
According to principle of least square method, the Generalized Least Square solution of the matrix B of composition is
B=(XTX)-1XTZ
The parameter A of the Gaussian function of fitting is found out further according to (II) formula1、B1And C1, can similarly find out the Gaussian function of fitting
Several parameter A2、B2And C2。
Further, vehicle part includes: license plate, logo, vehicle face, air inlet grill, the left front car light, the right side of front vehicle image
Front car light, left front fog lamp, right front fog lamp, left-hand mirror, right rear view mirror, windshield and skylight.
Further, benchmark is license plate, and other vehicle parts are relative angle relative to the relative position of license plate:Relative distance:W indicates that the width of license plate, H indicate
The height of license plate, license plate center point coordinate are (x0, y0), the center point coordinate of other vehicle parts is (x1, y1)。
It is another aspect of this invention to provide that a kind of vehicle part detection system based on relative position is provided, including from
Line training module and on-line checking module,
Off-line training module includes:
First submodule extracts the area-of-interest of sample vehicle part, selectes one for acquiring vehicle sample image
On the basis of the area-of-interest of sample vehicle part, the phase between the area-of-interest and benchmark of other sample vehicle parts is recorded
To position；
Second submodule is obtained prior information, is updated using prior information for establishing Gauss model using relative position
Thus raw score of the sample area-of-interest in fast convolution neural network obtains trained relative position network；
On-line checking module includes:
Vehicle image is inputted, the area-of-interest of vehicle part is extracted, area-of-interest input relative position network is obtained
To the score of area-of-interest, the area-of-interest of highest scoring is vehicle part target area.
Further, second submodule includes:
Gauss model module is established, for establishing Gauss model using relative position, obtains prior information, relative position packet
It includes: relative angle and relative distance:
Wherein, Δ1(x) prior information of the relative angle of the area-of-interest of vehicle part, Δ are indicated2(x) vehicle is indicated
The prior information of the relative distance of the area-of-interest of component, wherein Aj、Bj、CjParameter for the Gaussian function for needing to be fitted, L1
(x) relative angle of the area-of-interest of vehicle part, L are indicated2(x) relative angle of the area-of-interest of vehicle part is indicated
Degree, x indicate area-of-interest；
Update to obtain sub-module: for updating area-of-interest obtaining in fast convolution neural network using prior information
Point:
P*=arg max (Δ (x) s (x))
Pi *Indicate the score of the area-of-interest of vehicle part, Δ (x) indicates the opposite of the area-of-interest of vehicle part
The prior information of position, s (x) indicate that raw score of the area-of-interest of vehicle part in fast convolution neural network, α are
The weight of relative angle prior information, β are the weight of the prior information of relative distance.
Further, the fitting of the parameter of Gaussian function includes:
As j=1,
Equipped with one group of experimental data (xi, yi) (i=1,2,3 ... n), xiIt is one of the area-of-interest of vehicle part
Relative angle, yiFor the prior information of the relative position of the area-of-interest of vehicle part, then formula (I) can be written as:
Above formula both sides take natural logrithm, turn to
It enables
Consider total Test data, then indicate in the matrix form are as follows:
It is abbreviated as
Z=XB
According to principle of least square method, the Generalized Least Square solution of the matrix B of composition is
B=(XTX)-1XTZ
The parameter A of the Gaussian function of fitting is found out further according to (II) formula1、B1And C1, can similarly find out the Gaussian function of fitting
Several parameter A2、B2And C2。
Further, vehicle part includes: license plate, logo, vehicle face, air inlet grill, the left front car light, the right side of front vehicle image
Front car light, left front fog lamp, right front fog lamp, left-hand mirror, right rear view mirror, windshield and skylight.
Further, benchmark is license plate, the relative position information accuracy rate highest obtained on the basis of license plate, other vehicles
Component is relative angle relative to the relative position of license plate:Relative distance:W indicates that the width of license plate, H indicate that the height of license plate, license plate center point coordinate are (x0,
y0), the center point coordinate of other vehicle parts is (x1, y1)。
In general, through the invention it is contemplated above technical scheme is compared with the prior art, can obtain down and show
Beneficial effect:
In training relative position network, the relative position based on vehicle part obtains prior information, more using prior information
New score, effectively increases the precision of relative position network, to improve the accuracy rate of the target area detected；Based on vehicle
The relative position information of component carries out vehicle part detection and significantly improves the reliability of vehicle part detection, further enhances
The accuracy of vehicle part detection.
Detailed description of the invention
Fig. 1 is a kind of flow chart of vehicle part detection method based on relative position；
Fig. 2 is the schematic diagram of vehicle part.
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.
As shown in Figure 1, a kind of vehicle part detection method based on relative position, including off-line training and on-line checking portion
Point,
Wherein, off-line training part includes:
(1) vehicle sample image is acquired, the area-of-interest of sample vehicle part is extracted, selectes a sample vehicle part
Area-of-interest on the basis of, record the relative position between the area-of-interest and benchmark of other sample vehicle parts；
(2) Gauss model is established using relative position, obtains prior information, utilize prior information more new samples region of interest
Thus raw score of the domain in fast convolution neural network obtains trained relative position network；
On-line checking part includes:
Vehicle image is inputted, the area-of-interest of vehicle part is extracted, area-of-interest input relative position network is obtained
To the score of area-of-interest, the area-of-interest of highest scoring is vehicle part target area.
Further, as shown in Fig. 2, vehicle part includes: A: vehicle face；B: license plate；C: logo；D: air inlet grill；E: it keeps out the wind
Glass；F: left front car light；G: right front car light；H: skylight；I: left front fog lamp；J: right front fog lamp；K: left-hand mirror；L: right backsight
Mirror.
Further, the specific implementation of step (2) are as follows:
The relative position (2-1) includes: relative angle and relative distance, establishes Gauss model using relative position, obtains elder generation
Test information:
Wherein, Δ1(x) prior information of the relative angle of the area-of-interest of vehicle part, Δ are indicated2(x) vehicle is indicated
The prior information of the relative distance of the area-of-interest of component, wherein Aj、Bj、CjParameter for the Gaussian function for needing to be fitted, L1
(x) relative angle of the area-of-interest of vehicle part, L are indicated2(x) relative angle of the area-of-interest of vehicle part is indicated
Degree, x indicate area-of-interest；
(2-2) updates score of the area-of-interest in fast convolution neural network using prior information:
P*=arg max (Δ (x) s (x))
Pi *Indicate the score of the area-of-interest of vehicle part, Δ (x) indicates the opposite of the area-of-interest of vehicle part
The prior information of position, s (x) indicate that raw score of the area-of-interest of vehicle part in fast convolution neural network, α are
The weight of relative angle prior information, β are the weight of the prior information of relative distance.
Further, the fitting of the parameter of Gaussian function includes:
As j=1,
Equipped with one group of experimental data (xi, yi) (i=1,2,3 ... n), xiIt is one of the area-of-interest of vehicle part
Relative angle, yiFor the prior information of the relative position of the area-of-interest of vehicle part, then formula (I) can be written as:
Above formula both sides take natural logrithm, turn to
It enables
Consider total Test data, then indicate in the matrix form are as follows:
It is abbreviated as
Z=XB
According to principle of least square method, the Generalized Least Square solution of the matrix B of composition is
B=(XTX)-1XTZ
The parameter A of the Gaussian function of fitting is found out further according to (II) formula1、B1And C1, can similarly find out the Gaussian function of fitting
Several parameter A2、B2And C2。
Further, the relative position of vehicle part is angle:Relative distance:W indicates that the width of license plate, H indicate that the height of license plate, license plate center point coordinate are
(x0, y0), the center point coordinate of vehicle part is (x1, y1)。
Further, vehicle image is acquired, extracts the specific implementation method of the area-of-interest of vehicle part are as follows: by vehicle
Image input candidate region network obtains candidate region, and candidate region input region of interest network is obtained area-of-interest.
Preferably, the vehicle part candidate region generating mode of the embodiment of the present invention, which uses, includes the big of complete vehicle image
Mean aspect ratio and average area under sample statistics data, substantially increase the accuracy of candidate frame；
Specifically, extracting network using the trained primitive character of ImageNet class library extracts image spy to vehicle image
Sign, obtains characteristic image；To each position of characteristic image, 9 possible candidate regions are considered, three kinds of areas { 2,4,5 }, three
It plants length-width ratio { 3: 1,3: 2,3: 3 }, wherein the calculation of area and ratio are as follows:
Before generating candidate region, each side of characteristic image is zoomed in the range of 600~1000 at equal pace.
If the width of characteristic image and high respectively w and h, scaling formula are as follows:
Min=min (w, h)
Max=max (w, h)
If
If
The width of the candidate region of w ' expression vehicle part, the height of the candidate region of h ' expression vehicle part, scale1 are indicated
Characteristic image coefficient of reduction, scale2 indicate characteristic image amplification coefficient.
Further, candidate region network is any in Area generation network, selective search device and edge frame model
It is a kind of.
Further, candidate region network is Area generation network.
Further, area-of-interest network is fast convolution neural network, fast target detection convolutional neural networks, BP
Any one in neural network and convolutional neural networks.
Further, area-of-interest network is fast convolution neural network.
Preferably, the training of fast convolution neural network Fast R-CNN includes:
Step 1: micro- using model initialization RPN (Area generation network) network parameter of the pre-training on ImageNet
Adjust RPN network；
Step 2: candidate region training Fast R-CNN network is extracted using RPN network in step 1, also on ImageNet
The model initialization of the pre-training network parameter (it now appear that two networks are relatively independent)；
Step 3: reinitializing RPN using the Fast R-CNN network of step 2, fixed convolutional layer is finely adjusted, finely tunes
RPN network；
Step 4: the convolutional layer of Fast R-CNN in fixing step 2, the candidate region pair extracted using RPN in step 3
Fast R-CNN network is finely adjusted.
It is another aspect of this invention to provide that a kind of vehicle part detection system based on relative position is provided, including from
Line training module and on-line checking module,
Off-line training module includes:
First submodule extracts the area-of-interest of sample vehicle part, selectes one for acquiring vehicle sample image
On the basis of the area-of-interest of sample vehicle part, the phase between the area-of-interest and benchmark of other sample vehicle parts is recorded
To position；
Second submodule is obtained prior information, is updated using prior information for establishing Gauss model using relative position
Thus raw score of the sample area-of-interest in fast convolution neural network obtains trained relative position network；
On-line checking module includes:
Vehicle image is inputted, the area-of-interest of vehicle part is extracted, area-of-interest input relative position network is obtained
To the score of area-of-interest, the area-of-interest of highest scoring is vehicle part target area.
Further, second submodule includes:
Gauss model module is established, for establishing Gauss model using relative position, obtains prior information, relative position packet
It includes: relative angle and relative distance:
Wherein, Δ1(x) prior information of the relative angle of the area-of-interest of vehicle part, Δ are indicated2(x) vehicle is indicated
The prior information of the relative distance of the area-of-interest of component, wherein Aj、Bj、CjParameter for the Gaussian function for needing to be fitted, L1
(x) relative angle of the area-of-interest of vehicle part, L are indicated2(x) relative angle of the area-of-interest of vehicle part is indicated
Degree, x indicate area-of-interest；
Update to obtain sub-module: for updating area-of-interest obtaining in fast convolution neural network using prior information
Point:
P*=arg max (Δ (x) s (x))
Pi *Indicate the score of the area-of-interest of vehicle part, Δ (x) indicates the opposite of the area-of-interest of vehicle part
The prior information of position, s (x) indicate that raw score of the area-of-interest of vehicle part in fast convolution neural network, α are
The weight of relative angle prior information, β are the weight of the prior information of relative distance.
Further, the fitting of the parameter of Gaussian function includes:
As j=1,
Equipped with one group of experimental data (xi, yi) (i=1,2,3 ... n), xiIt is one of the area-of-interest of vehicle part
Relative angle, yiFor the prior information of the relative position of the area-of-interest of vehicle part, then formula (I) can be written as:
Above formula both sides take natural logrithm, turn to
It enables
Consider total Test data, then indicate in the matrix form are as follows:
It is abbreviated as
Z=XB
According to principle of least square method, the Generalized Least Square solution of the matrix B of composition is
B=(XTX)-1XTZ
The parameter A of the Gaussian function of fitting is found out further according to (II) formula1、B1And C1, can similarly find out the Gaussian function of fitting
Several parameter A2、B2And C2。
Further, vehicle part includes: license plate, logo, vehicle face, air inlet grill, the left front car light, the right side of front vehicle image
Front car light, left front fog lamp, right front fog lamp, left-hand mirror, right rear view mirror, windshield and skylight.
Further, benchmark is license plate, the relative position information accuracy rate highest obtained on the basis of license plate, other vehicles
Component is relative angle relative to the relative position of license plate:Relative distance:W indicates that the width of license plate, H indicate that the height of license plate, license plate center point coordinate are
(x0, y0), the center point coordinate of other vehicle parts is (x1, y1)。
Table 1 gives the testing result detected using the method for the present invention to 10000 vehicle pictures, for vehicle
The F1 value of the components such as face, license plate, air inlet grill, windshield, left front car light, right front car light, left-hand mirror, right rear view mirror detection is all
Greater than 0.9, wherein license plate, vehicle face, windshield are even more to be higher than 0.96, the knot of the method for far superior to traditional target detection
Fruit；And the F1 value of the detection of logo is greater than 0.82, left fog lamp, right fog lamp, skylight the F1 value of detection also reached 0.64, relatively
Traditional object detection method has greatly improved.
Table 1
| Accuracy | Recall rate | Harmonic average (F1) | |
| Vehicle face | 0.9850 | 0.9868 | 0.9859 |
| Logo | 0.8764 | 0.7777 | 0.8241 |
| Air inlet grill | 0.9351 | 0.9220 | 0.9285 |
| License plate | 0.9714 | 0.9620 | 0.9667 |
| Windshield | 0.9827 | 0.9846 | 0.9836 |
| Left front car light | 0.9562 | 0.9001 | 0.9273 |
| Right front car light | 0.9554 | 0.9046 | 0.9293 |
| Left fog lamp | 0.7322 | 0.5875 | 0.6519 |
| Right fog lamp | 0.7406 | 0.5693 | 0.6438 |
| Skylight | 0.6775 | 0.7364 | 0.7057 |
| Left-hand mirror | 0.9278 | 0.8738 | 0.9000 |
| Right rear view mirror | 0.9225 | 0.8842 | 0.9029 |
It can be seen that the relative position based on vehicle part obtains prior information in training relative position network, utilize
Prior information corrects scoring function, the precision of relative position network is effectively increased, to improve the target area detected
Accuracy rate；What relative position information based on vehicle part carried out that vehicle part detection significantly improves vehicle part detection can
By property, the accuracy of vehicle part detection is further enhanced.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should all include
Within protection scope of the present invention.
Claims (8)
1. a kind of vehicle part detection method based on relative position, which is characterized in that including off-line training and on-line checking portion
Point,
The off-line training part includes:
(1) vehicle sample image is acquired, the area-of-interest of sample vehicle part is extracted, selectes the sense of a sample vehicle part
On the basis of interest region, the relative position between the other area-of-interests and benchmark of sample vehicle part is recorded；
(2) Gauss model is established using relative position, obtains prior information, existed using prior information more new samples area-of-interest
Thus raw score in fast convolution neural network obtains trained relative position network；
The on-line checking part includes:
Vehicle image is inputted, the area-of-interest of vehicle part is extracted, area-of-interest input relative position network is felt
The score in interest region, the area-of-interest of highest scoring are vehicle part target area；
The specific implementation of the step (2) are as follows:
The relative position (2-1) includes: relative angle and relative distance, establishes Gauss model using relative position, obtains priori letter
Breath:
Wherein, Δ1(x) prior information of the relative angle of the area-of-interest of vehicle part, Δ are indicated2(x) vehicle part is indicated
Area-of-interest relative distance prior information, wherein Aj、Bj、CjParameter for the Gaussian function for needing to be fitted, L1(x)
Indicate the relative angle of the area-of-interest of vehicle part, L2(x) relative angle of the area-of-interest of vehicle part, x are indicated
Indicate area-of-interest；
(2-2) updates score of the area-of-interest in fast convolution neural network using prior information:
P*=argmax (Δ (x) s (x))
Pi *Indicate the score of the area-of-interest of vehicle part, Δ (x) indicates the relative position of the area-of-interest of vehicle part
Prior information, s (x) indicates raw score of the area-of-interest in fast convolution neural network of vehicle part, and α is opposite
The weight of angle prior information, β are the weight of the prior information of relative distance.
2. a kind of vehicle part detection method based on relative position as described in claim 1, which is characterized in that the Gauss
The fitting of the parameter of function includes:
As j=1,
Equipped with one group of experimental data (xi,yi), i=1,2,3 ... n, xiFor a relative angle of the area-of-interest of vehicle part
Degree, yiFor the prior information of the relative position of the area-of-interest of vehicle part, then formula (I) can be written as:
Above formula both sides take natural logrithm, turn to
It enables
Consider total Test data, then indicate in the matrix form are as follows:
It is abbreviated as
Z=XB
According to principle of least square method, the Generalized Least Square solution of the matrix B of composition is
B=(XTX)-1XTZ
The parameter A of the Gaussian function of fitting is found out further according to (II) formula1、B1And C1, can similarly find out the Gaussian function of fitting
Parameter A2、B2And C2。
3. a kind of vehicle part detection method based on relative position as described in claim 1, which is characterized in that the vehicle
Component includes: license plate, logo, vehicle face, air inlet grill, left front car light, the right front car light of front vehicle image, left front fog lamp, it is right before
Fog lamp, left-hand mirror, right rear view mirror, windshield and skylight.
4. a kind of vehicle part detection method based on relative position as claimed in claim 3, which is characterized in that the benchmark
For license plate, other vehicle parts are relative angle relative to the relative position of license plate:Relative distance:W indicates that the width of license plate, H indicate that the height of license plate, license plate center point coordinate are (x0,
y0), the center point coordinate of other vehicle parts is (x1,y1)。
5. a kind of vehicle part detection system based on relative position, which is characterized in that including off-line training module and online inspection
Module is surveyed,
The off-line training module includes:
First submodule extracts the area-of-interest of sample vehicle part, selectes a sample for acquiring vehicle sample image
On the basis of the area-of-interest of vehicle part, the opposite position between the other area-of-interests and benchmark of sample vehicle part is recorded
It sets；
Second submodule obtains prior information, utilizes prior information more new samples for establishing Gauss model using relative position
Thus raw score of the area-of-interest in fast convolution neural network obtains trained relative position network；
The on-line checking module includes:
Vehicle image is inputted, the area-of-interest of vehicle part is extracted, area-of-interest input relative position network is felt
The score in interest region, the area-of-interest of highest scoring are vehicle part target area；
The second submodule includes:
Gauss model module is established, for establishing Gauss model using relative position, obtains prior information, relative position includes:
Relative angle and relative distance:
Wherein, Δ1(x) prior information of the relative angle of the area-of-interest of vehicle part, Δ are indicated2(x) vehicle part is indicated
Area-of-interest relative distance prior information, wherein Aj、Bj、CjParameter for the Gaussian function for needing to be fitted, L1(x)
Indicate the relative angle of the area-of-interest of vehicle part, L2(x) relative angle of the area-of-interest of vehicle part, x are indicated
Indicate area-of-interest；
Update to obtain sub-module: for updating score of the area-of-interest in fast convolution neural network using prior information:
P*=argmax (Δ (x) s (x))
Pi *Indicate the score of the area-of-interest of vehicle part, Δ (x) indicates the relative position of the area-of-interest of vehicle part
Prior information, s (x) indicates raw score of the area-of-interest in fast convolution neural network of vehicle part, and α is opposite
The weight of angle prior information, β are the weight of the prior information of relative distance.
6. a kind of vehicle part detection system based on relative position as claimed in claim 5, which is characterized in that the Gauss
The fitting of the parameter of function includes:
As j=1,
Equipped with one group of experimental data (xi,yi), i=1,2,3 ... n, xiFor a relative angle of the area-of-interest of vehicle part
Degree, yiFor the prior information of the relative position of the area-of-interest of vehicle part, then formula (I) can be written as:
Above formula both sides take natural logrithm, turn to
It enables
Consider total Test data, then indicate in the matrix form are as follows:
It is abbreviated as
Z=XB
According to principle of least square method, the Generalized Least Square solution of the matrix B of composition is
B=(XTX)-1XTZ
The parameter A of the Gaussian function of fitting is found out further according to (II) formula1、B1And C1, can similarly find out the Gaussian function of fitting
Parameter A2、B2And C2。
7. a kind of vehicle part detection system based on relative position as claimed in claim 5, which is characterized in that the vehicle
Component includes: license plate, logo, vehicle face, air inlet grill, left front car light, the right front car light of front vehicle image, left front fog lamp, it is right before
Fog lamp, left-hand mirror, right rear view mirror, windshield and skylight.
8. a kind of vehicle part detection system based on relative position as claimed in claim 7, which is characterized in that the benchmark
For license plate, other vehicle parts are relative angle relative to the relative position of license plate:Relative distance:W indicates that the width of license plate, H indicate that the height of license plate, license plate center point coordinate are (x0,
y0), the center point coordinate of other vehicle parts is (x1,y1)。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710234936.4A CN107016390B (en) | 2017-04-11 | 2017-04-11 | A kind of vehicle part detection method and system based on relative position |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710234936.4A CN107016390B (en) | 2017-04-11 | 2017-04-11 | A kind of vehicle part detection method and system based on relative position |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107016390A CN107016390A (en) | 2017-08-04 |
| CN107016390B true CN107016390B (en) | 2019-11-12 |
Family
ID=59445501
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710234936.4A Active CN107016390B (en) | 2017-04-11 | 2017-04-11 | A kind of vehicle part detection method and system based on relative position |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107016390B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108256498A (en) * | 2018-02-01 | 2018-07-06 | 上海海事大学 | A kind of non power driven vehicle object detection method based on EdgeBoxes and FastR-CNN |
| CN108491827B (en) * | 2018-04-13 | 2020-04-10 | 腾讯科技（深圳）有限公司 | Vehicle detection method and device and storage medium |
| TWI699705B (en) * | 2018-08-06 | 2020-07-21 | 拓連科技股份有限公司 | Position determination methods and systems for vehicle, and related computer program prodcuts |
| CN109741398A (en) * | 2018-12-28 | 2019-05-10 | 同方威视技术股份有限公司 | Image-recognizing method, device and computer storage medium |
| CN113256625A (en) * | 2021-06-29 | 2021-08-13 | 北京天星博迈迪医疗器械有限公司 | Electronic equipment and recognition device |
| CN114821478B (en) * | 2022-05-05 | 2023-01-13 | 北京容联易通信息技术有限公司 | Process flow detection method and system based on video intelligent analysis |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103488973A (en) * | 2013-09-12 | 2014-01-01 | 上海依图网络科技有限公司 | Method and system for recognizing vehicle brand based on image |
| CN105160330A (en) * | 2015-09-17 | 2015-12-16 | 中国地质大学（武汉） | Vehicle logo recognition method and vehicle logo recognition system |
| CN105740910A (en) * | 2016-02-02 | 2016-07-06 | 北京格灵深瞳信息技术有限公司 | Vehicle object detection method and device |
-
2017
- 2017-04-11 CN CN201710234936.4A patent/CN107016390B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103488973A (en) * | 2013-09-12 | 2014-01-01 | 上海依图网络科技有限公司 | Method and system for recognizing vehicle brand based on image |
| CN105160330A (en) * | 2015-09-17 | 2015-12-16 | 中国地质大学（武汉） | Vehicle logo recognition method and vehicle logo recognition system |
| CN105740910A (en) * | 2016-02-02 | 2016-07-06 | 北京格灵深瞳信息技术有限公司 | Vehicle object detection method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107016390A (en) | 2017-08-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107016390B (en) | A kind of vehicle part detection method and system based on relative position | |
| WO2019196130A1 (en) | Classifier training method and device for vehicle-mounted thermal imaging pedestrian detection | |
| CN102509098B (en) | Fisheye image vehicle identification method | |
| CN110399856B (en) | Feature extraction network training method, image processing method, device and equipment | |
| Pavlic et al. | Classification of images in fog and fog-free scenes for use in vehicles | |
| DE102012218390B4 (en) | Optimizing the detection of objects in images | |
| CN110348445A (en) | A kind of example dividing method merging empty convolution sum marginal information | |
| CN107767405A (en) | A kind of nuclear phase for merging convolutional neural networks closes filtered target tracking | |
| CN111027481B (en) | Behavior analysis method and device based on human body key point detection | |
| TW201135680A (en) | Vehicle tracking system and tracking method thereof | |
| CN102096821A (en) | Number plate identification method under strong interference environment on basis of complex network theory | |
| CN111489330B (en) | Weak and small target detection method based on multi-source information fusion | |
| CN104021552B (en) | Multi-objective particle swarm parameter optimization method based on graph segmentation process | |
| CN108765327A (en) | A kind of image rain removing method based on the depth of field and sparse coding | |
| CN111461039A (en) | Landmark identification method based on multi-scale feature fusion | |
| CN113792606B (en) | Low-cost self-supervision pedestrian re-identification model construction method based on multi-target tracking | |
| CN113111722A (en) | Automatic driving target identification method based on improved Mask R-CNN | |
| Kim et al. | Effective traffic lights recognition method for real time driving assistance systemin the daytime | |
| CN109657599B (en) | Picture identification method of distance-adaptive vehicle appearance part | |
| CN104112138A (en) | Object color classification method and device | |
| Juneja et al. | A systematic review on foggy datasets: Applications and challenges | |
| Tu et al. | A biologically inspired vision-based approach for detecting multiple moving objects in complex outdoor scenes | |
| CN112347967B (en) | Pedestrian detection method fusing motion information in complex scene | |
| Ren et al. | Research on infrared small target segmentation algorithm based on improved mask R-CNN | |
| TW202020724A (en) | Image adaptive feature extraction method and application thereof integrating each adaptive feature block to form an adaptive feature image |
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 |