CN104751190A - Vehicle part positioning method for vehicle fine recognition - Google Patents

Abstract

The invention discloses a vehicle part positioning method for vehicle fine recognition. A precise vehicle part positioning result is obtained by model training and part detecting. The model training is characterized in that HOG feature extraction is performed on the training data sets of labeled vehicles and vehicle parts, SVM classifier training is used to obtain model parameters, and a model which can be used for vehicle part detecting is obtained. The part detecting is characterized in that the feature pyramid of a to-be-tested image is extracted, the obtained model is used to perform convolution operation on the feature pyramid to obtain a convolution response value, the convolution response value is combined with deformation constraint to obtain the rough positioning of vehicle parts, and rigid body feature, position constraint and symmetry constraint are used to achieve precise positioning of the vehicle parts. By the method, precise positioning of the vehicle parts in the image can be achieved, problems in the background technology are solved effectively, and foundation work of vehicle fine recognition is achieved.

Description

A kind of vehicle part localization method towards the meticulous identification of vehicle

Technical field

The invention belongs to mode identification technology, particularly relate to a kind of vehicle part localization method towards the meticulous identification of vehicle.

Background technology

Continue the soaring and generally erection of road high-definition monitoring camera and widely using of individual smart mobile phone along with vehicle population, all types of car picture number that relates to becomes geometric series to double.Due in the past simple utilize artificial from picture the information factor data amount such as identification vehicle model huge, and the identification capability of people limited and be difficult to realize, therefore how to utilize computer technology to carry out automatic Identification to vehicle brand and model and developed into a large important research direction, it will play important support effect to the field such as intelligent transportation, social public security.

The automatic Study of recognition of vehicle can be divided into two levels: ground floor carries out coarseness sort research to vehicle, realizes the judgement of vehicle fundamental type, as vehicle is divided into car, SUV, bus, truck etc.; The second layer is then carry out fine grit classification research to vehicle, attempts the identification realized brand belonging to vehicle and model, as identified Audi Q7 or BMW is X5.The range of application of vehicle coarseness classification mainly concentrates on the occasion that highway toll etc. only pays close attention to type of vehicle, detects the system of checking etc. require cannot be suitable in the application of concrete type information in vehicle management system, public security.Vehicle fine grit classification not only accurately can identify brand in picture and model, can also find fake license plate vehicle etc., have the scope of application and more far-reaching realistic meaning widely compared with coarseness classification by the register information of comparison vehicle and car plate.

Up to now, the method of existing vehicle cab recognition mainly concentrates in the sort research of vehicle coarseness, as utilized Edge Feature Points in vehicle outer contour or [document 2] to classify in [document 1], part research [document 3,4] is also had to classify to the vehicle in image or video based on the color of vehicle or whole geometry feature.More than study the overall significant characteristics mainly utilizing target, but often have similar global characteristics between target type in fine grit classification research, as similar profile, identical ingredient etc., dissimilar only exists trickle difference in appearance.

The existing model recognizing method based on coarseness classification utilizes the global feature difference of vehicle to classify, and only can identify fundamental type (as car, passenger vehicle, truck etc.), be of limited application.Therefore, need research vehicle fine grit classification method, by positioned vehicle building block and the nuance excavating each parts reaches the object of vehicle brand and type identifier.And in order to meet the needs of fine grit classification, in the urgent need to a kind of new method, vehicle part is accurately located.

[document 1] Jolly, M-PD, S.Lakshmanan, and A.K.Jain, Vehicle Segmentation andClassification Using Deformable Templates, Pattern Analysis and Machine Intelligence, IEEE Transactions on, 1996.18 (3): p.293-308.

[document 2] Ma, X, W.E.L.Grimson, Edge-based rich representation for vehicle classificatio, in Computer Vision, IEEE International Conference on, 2005.p.1185-1192.

[document 3] Kafai, M, B.Bhanu, Dynamic Bayesian Networks for Vehicle Classification inVideo.Industrial Informatics, IEEE Transactions on, 2012.8 (1).

[document 4] Moussa, G.S, Vehicle Type Classification with Geometric and AppearanceAttributes.International Journal of Civil, Architectural Science and Engineering, 2014.8 (3): p.273-278.

Summary of the invention

In order to solve above-mentioned technical matters, the present invention proposes a kind of vehicle part localization method towards the meticulous identification of vehicle.

The technical solution adopted in the present invention is: a kind of vehicle part localization method towards the meticulous identification of vehicle, is characterized in that: model training and parts accurately locate two benches;

In the described model training stage, its specific implementation comprises following sub-step:

Step 1.1: training dataset prepares, and training dataset comprises positive sample data and negative sample data; Positive sample data is the view data having marked vehicle housing and parts housing, and negative sample data are the view data beyond positive sample data;

Step 1.2: data set is classified, uses K average sorter, separates the data under different visual angles according to the position of parts and the length breadth ratio of vehicle;

Step 1.3: data set pre-service, according to position and the weights relation of parts mark, obtains minimum spanning tree model;

Step 1.4: initialization obtains vehicle detection model, this initialization procedure is that establishment root wave filter and parts wave filter initial value are zero, other are for the pre-set vehicle detection model of the parameter detected, a detection model is made up of a root wave filter and several parts wave filters, and its concrete manifestation form is (F ₀, P ₁..., P _n, b), wherein, F ₀what represent is root wave filter, P _iwhat represent is i-th parts wave filter, and b is a biased coefficient; Each parts wave filter has a tlv triple (F _i, v _i, d _i) represent, wherein F _ithe value of i-th parts wave filter, v _ibe one and specify the bivector of parts wave filter relative to root filter location, d _irepresent the displacement loss of parts wave filter; When there being multiple visual angle, the form that embodies of vehicle detection model parameter is:

β＝(F′ ₀，...，F′ _n，d ₁，...，d _n，b) (1)；

Wherein F ' _irepresent the vehicle detection model under i-th visual angle;

The concrete form of root wave filter and parts wave filter is the convolution mask of N × M size, and wherein N and M determines by the estimation process of training dataset; With the vehicle detection model training described in training data set pair, obtain root wave filter and parts wave filter; The process of its training is use formula (2) to utilize gradient descent method to the solution procedure of β:

$L_D\left(\mathrm{\β}\right)=\frac{1}{2}{\left|\right|\mathrm{\β}\left|\right|}^2+C{\mathrm{\Σ}}_{i=1}^n\max (\mathrm{0,1}-y_i{f}_{\mathrm{\β}}\left(x_i\right))---\left(2\right);$

Wherein D=(<x ₁, y ₁> ..., <x _n, y _n>) be a series of nominal data, β is model parameter, max (0,1-y _if _β(x _i) be the loss function of standard, C controls the relative weighting of regular terms;

Parts wave filter employs minimum spanning tree model and carries out position constraint, and comprehensive root wave filter and parts wave filter, obtain vehicle part location model;

The accurate positioning stage of described parts, its specific implementation comprises following sub-step:

Step 2.1: the vehicle part location model described in reading;

Step 2.2: for needing the picture carrying out vehicle part location, extracts HOG feature pyramid, and HOG feature pyramid is and extracts HOG feature to picture under different resolution;

Step 2.3: utilize root wave filter and the pyramidal designated layer L of feature to carry out convolution, obtain convolution response, the part that score is high is vehicle region;

Step 2.4: utilize parts wave filter and feature pyramid L × 2 layer to carry out convolution, obtain convolution response, the vehicle region obtained in integrating step 2.3, add displacement loss constraint, the place that score is high is component area; The concrete formula calculated the score is:

Wherein:

(dx _i，dy _i)＝(x _i，y _i)-(2(x ₀，y ₀)+v _i) (4)；

In formula (3), p _i=(x _i, y _i, l _i) specify the number of plies of i-th wave filter in feature pyramid and corresponding position, d _irepresent the displacement loss of parts wave filter; Formula (4) specifies the displacement loss of the relative root wave filter of parts wave filter, (x _i, y _i) represent the position detecting parts, (x ₀, y ₀) represent the position of parts in model, v _ibe one and specify the bivector of parts wave filter relative to root filter location; Formula (5) represents deformation characteristics;

Step 2.5: the position relationship characteristic and the symmetric properties that utilize vehicle, to the parts point of addition detected and symmetric constraints, obtain vehicle part and accurately locate.

As preferably, the specific implementation of step 1.4 comprises following sub-step:

Step 1.4.1: initialization vehicle detection model, utilizes training dataset to carry out the model training of the first step, obtains a master pattern;

Step 1.4.2: produce a symmetrical model, carry out the model training of second step, obtaining can for the symmetrical model carrying out detecting of left and right vehicle wheel;

Step 1.4.3: the model obtained in combining step 1.4.1 and step 1.4.2, carries out the model training of the 3rd step, and obtain a complete vehicle detection model that can carry out symmetrical detection to vehicle, this model obtains root wave filter;

Step 1.4.4: utilize the vehicle detection model obtained in step 1.4.3, align sample data and detect, align sample data and again demarcate;

Step 1.4.5: use the vehicle detection model obtained in step 1.4.3, by the parts data marked, carries out the model training of the 4th step, obtains parts detection model;

Step 1.4.6: add minimum spanning tree constraint to parts detection model, adjust component locations, for the further training component detection model of parts, this model obtains parts wave filter;

Step 1.4.7: the parts detection model obtained in the vehicle detection model obtained in combining step 1.4.3 and step 1.4.6, by root wave filter and parts filter synthesis, obtains vehicle part location model.

The present invention can realize carrying out parts to the vehicle in picture and accurately locate, thus effective problem solving background technology and mention, be the element task of the meticulous identification of vehicle.

Accompanying drawing explanation

Fig. 1: the model training phase flow figure of the embodiment of the present invention;

Fig. 2: the accurate positioning stage process flow diagram of parts of the embodiment of the present invention.

Embodiment

Understand for the ease of those of ordinary skill in the art and implement the present invention, below in conjunction with drawings and Examples, the present invention is described in further detail, should be appreciated that exemplifying embodiment described herein is only for instruction and explanation of the present invention, is not intended to limit the present invention.

A kind of vehicle part localization method model training towards the meticulous identification of vehicle provided by the invention and parts accurately locate two benches;

Ask for an interview Fig. 1, in the model training stage of the present embodiment, its specific implementation comprises following sub-step:

Step 1.1: training dataset prepares, and training dataset comprises positive sample data and negative sample data; Positive sample data is the view data having marked vehicle housing and parts housing, and negative sample data are the view data beyond positive sample data;

Step 1.2: data set is classified, uses K average sorter, separates the data under different visual angles according to the position of parts and the length breadth ratio of vehicle;

Step 1.3: data set pre-service, according to position and the weights relation of parts mark, obtains minimum spanning tree model;

Step 1.4: initialization obtains vehicle detection model, this initialization procedure is that establishment root wave filter and parts wave filter initial value are zero, other are for the pre-set vehicle detection model of the parameter detected, a detection model is made up of a root wave filter and several parts wave filters, and its concrete manifestation form is (F ₀, P ₁..., P _n, b), wherein, F ₀what represent is root wave filter, P _iwhat represent is i-th parts wave filter, and b is a biased coefficient; Each parts wave filter has a tlv triple (F _i, v _i, d _i) represent, wherein F _ithe value of i-th parts wave filter, v _ibe one and specify the bivector of parts wave filter relative to root filter location, d _irepresent the displacement loss of parts wave filter; When there being multiple visual angle, the form that embodies of vehicle detection model parameter is:

β＝(F′ ₀，...，F′ _n，d ₁，...，d _n，b) (1)；

Wherein F ' _irepresent the vehicle detection model under i-th visual angle;

The concrete form of root wave filter and parts wave filter is the convolution mask of N × M size, and wherein N and M determines by the estimation process of training dataset; With the vehicle detection model training described in training data set pair, obtain root wave filter and parts wave filter; The process of its training is use formula (2) to utilize gradient descent method to the solution procedure of β:

$L_D\left(\mathrm{\&beta;}\right)=\frac{1}{2}{\left|\right|\mathrm{\&beta;}\left|\right|}^2+C{\mathrm{\&Sigma;}}_{i=1}^n\max (\mathrm{0,1}-y_i{f}_{\mathrm{\&beta;}}\left(x_i\right))---\left(2\right);$

Wherein D=(<x ₁, y ₁> ..., <x _n, y _n>) be a series of nominal data, β is model parameter, max (0,1-y _if _β(x _i) be the loss function of standard, C controls the relative weighting of regular terms;

Parts wave filter employs minimum spanning tree model and carries out position constraint, and comprehensive root wave filter and parts wave filter, obtain vehicle part location model;

Specific implementation comprises following sub-step:

Step 1.4.1: initialization vehicle detection model, utilizes training dataset to carry out the model training of the first step, obtains a master pattern;

Step 1.4.2: produce a symmetrical model, carry out the model training of second step, obtaining can for the symmetrical model carrying out detecting of left and right vehicle wheel;

Step 1.4.3: the model obtained in combining step 1.4.1 and step 1.4.2, carries out the model training of the 3rd step, and obtain a complete vehicle detection model that can carry out symmetrical detection to vehicle, this model obtains root wave filter;

Step 1.4.4: utilize the vehicle detection model obtained in step 1.4.3, align sample data and detect, align sample data and again demarcate;

Step 1.4.5: use the vehicle detection model obtained in step 1.4.3, by the parts data marked, carries out the model training of the 4th step, obtains parts detection model;

Step 1.4.6: add minimum spanning tree constraint to parts detection model, adjust component locations, for the further training component detection model of parts, this model obtains parts wave filter;

Step 1.4.7: the parts detection model obtained in the vehicle detection model obtained in combining step 1.4.3 and step 1.4.6, by root wave filter and parts filter synthesis, obtains vehicle part location model.

Ask for an interview Fig. 2, the accurate positioning stage of parts of the present embodiment, its specific implementation comprises following sub-step:

Step 2.1: the vehicle part location model described in reading;

Step 2.2: for needing the picture carrying out vehicle part location, extracts HOG feature pyramid, and HOG feature pyramid is and extracts HOG feature to picture under different resolution;

Step 2.3: utilize root wave filter and the pyramidal designated layer L of feature to carry out convolution, obtain convolution response, the part that score is high is vehicle region;

Step 2.4: utilize parts wave filter and feature pyramid L × 2 layer to carry out convolution, obtain convolution response, the vehicle region obtained in integrating step 2.3, add displacement loss constraint, the place that score is high is component area; The concrete formula calculated the score is:

Wherein:

(dx _i，dy _i)＝(x _i，y _i)-(2(x ₀，y ₀)+v _i) (4)；

In formula (3), p _i=(x _i, y _i, l _i) specify the number of plies of i-th wave filter in feature pyramid and corresponding position, d _irepresent the displacement loss of parts wave filter; Formula (4) specifies the displacement loss of the relative root wave filter of parts wave filter, (x _i, y _i) represent the position detecting parts, (x ₀, y ₀) represent the position of parts in model, v _ibe one and specify the bivector of parts wave filter relative to root filter location; Formula (5) represents deformation characteristics;

Step 2.5: the position relationship characteristic and the symmetric properties that utilize vehicle, to the parts point of addition detected and symmetric constraints, obtain vehicle part and accurately locate.

Should be understood that, the part that this instructions does not elaborate all belongs to prior art.

Should be understood that; the above-mentioned description for preferred embodiment is comparatively detailed; therefore the restriction to scope of patent protection of the present invention can not be thought; those of ordinary skill in the art is under enlightenment of the present invention; do not departing under the ambit that the claims in the present invention protect; can also make and replacing or distortion, all fall within protection scope of the present invention, request protection domain of the present invention should be as the criterion with claims.

Claims (2)

1. towards a vehicle part localization method for the meticulous identification of vehicle, it is characterized in that: model training and parts accurately locate two benches;

In the described model training stage, its specific implementation comprises following sub-step:

Step 1.1: training dataset prepares, and training dataset comprises positive sample data and negative sample data; Positive sample data is the view data having marked vehicle housing and parts housing, and negative sample data are the view data beyond positive sample data;

Step 1.2: data set is classified, uses K average sorter, separates the data under different visual angles according to the position of parts and the length breadth ratio of vehicle;

Step 1.3: data set pre-service, according to position and the weights relation of parts mark, obtains minimum spanning tree model;

Step 1.4: initialization obtains vehicle detection model, this initialization procedure is that establishment root wave filter and parts wave filter initial value are zero, other are for the pre-set vehicle detection model of the parameter detected, a detection model is made up of a root wave filter and several parts wave filters, and its concrete manifestation form is (F ₀, P ₁..., P _n, b), wherein, F ₀what represent is root wave filter, P _iwhat represent is i-th parts wave filter, and b is a biased coefficient; Each parts wave filter has a tlv triple (F _i, v _i, d _i) represent, wherein F _ithe value of i-th parts wave filter, v _ibe one and specify the bivector of parts wave filter relative to root filter location, d _irepresent the displacement loss of parts wave filter; When there being multiple visual angle, the form that embodies of vehicle detection model parameter is:

β＝(F′ ₀，…，F′ _n，d ₁，…，d _n，b) (1)；

Wherein F ' _irepresent the vehicle detection model under i-th visual angle;

The concrete form of root wave filter and parts wave filter is the convolution mask of N × M size, and wherein N and M determines by the estimation process of training dataset; With the vehicle detection model training described in training data set pair, obtain root wave filter and parts wave filter; The process of its training is use formula (2) to utilize gradient descent method to the solution procedure of β:

$L_D\left(\mathrm{\&beta;}\right)=\frac{1}{2}{\left|\right|\mathrm{\&beta;}\left|\right|}^2+C{\mathrm{\&Sigma;}}_{i=1}^n\max (0.1-y_i{f}_{\mathrm{\&beta;}}\left(x_i\right))---\left(2\right);$

Wherein D=(<x ₁, y ₁> ..., <x _n, y _n>) be a series of nominal data, β is model parameter, max (0,1-y _if _β(x _i) be the loss function of standard, C controls the relative weighting of regular terms;

Parts wave filter employs minimum spanning tree model and carries out position constraint, and comprehensive root wave filter and parts wave filter, obtain vehicle part location model;

The accurate positioning stage of described parts, its specific implementation comprises following sub-step:

Step 2.1: the vehicle part location model described in reading;

Step 2.2: for needing the picture carrying out vehicle part location, extracts HOG feature pyramid, and HOG feature pyramid is and extracts HOG feature to picture under different resolution;

Step 2.3: utilize root wave filter and the pyramidal designated layer L of feature to carry out convolution, obtain convolution response, the part that score is high is vehicle region;

Step 2.4: utilize parts wave filter and feature pyramid L × 2 layer to carry out convolution, obtain convolution response, the vehicle region obtained in integrating step 2.3, add displacement loss constraint, the place that score is high is component area; The concrete formula calculated the score is:

Wherein:

(dx _i，dy _i)＝(x _i，y _i)-(2(x ₀，y ₀)+v _i) (4)；

In formula (3), p _i=(x _i, y _i, l _i) specify the number of plies of i-th wave filter in feature pyramid and corresponding position, d _irepresent the displacement loss of parts wave filter; Formula (4) specifies the displacement loss of the relative root wave filter of parts wave filter, (x _i, y _i) represent the position detecting parts, (x ₀, y ₀) represent the position of parts in model, v _ibe one and specify the bivector of parts wave filter relative to root filter location; Formula (5) represents deformation characteristics;

Step 2.5: the position relationship characteristic and the symmetric properties that utilize vehicle, to the parts point of addition detected and symmetric constraints, obtain vehicle part and accurately locate.

2. the vehicle part localization method towards the meticulous identification of vehicle according to claim 1, is characterized in that: the specific implementation of step 1.4 comprises following sub-step:

Step 1.4.1: initialization vehicle detection model, utilizes training dataset to carry out the model training of the first step, obtains a master pattern;

Step 1.4.2: produce a symmetrical model, carry out the model training of second step, obtaining can for the symmetrical model carrying out detecting of left and right vehicle wheel;

Step 1.4.3: the model obtained in combining step 1.4.1 and step 1.4.2, carries out the model training of the 3rd step, and obtain a complete vehicle detection model that can carry out symmetrical detection to vehicle, this model obtains root wave filter;

Step 1.4.4: utilize the vehicle detection model obtained in step 1.4.3, align sample data and detect, align sample data and again demarcate;

Step 1.4.5: use the vehicle detection model obtained in step 1.4.3, by the parts data marked, carries out the model training of the 4th step, obtains parts detection model;

Step 1.4.6: add minimum spanning tree constraint to parts detection model, adjust component locations, for the further training component detection model of parts, this model obtains parts wave filter;

Step 1.4.7: the parts detection model obtained in the vehicle detection model obtained in combining step 1.4.3 and step 1.4.6, by root wave filter and parts filter synthesis, obtains vehicle part location model.