CN105335758A - Model identification method based on video Fisher vector descriptors - Google Patents

Abstract

The invention discloses a model identification method based on video Fisher vector descriptors. The method comprises the following steps: first of all, tacking a vehicle for training, tracking an image of each model in a video, and extracting an SIFT characteristic of each frame of the image of the model; then, performing Fisher vector coding calculation on all the SIFT characteristics of the model image; then performing PCA dimension reduction on obtained Fisher vector descriptors; then performing binarization on the descriptors after the dimension reduction to obtain the video Fisher vector descriptors of the model; performing SVM training on all the obtained descriptors to obtain an identification system with N model types; and for a vehicle video for testing, extracting the video Fisher vector descriptors of the vehicle video, introducing the video Fisher vector descriptors into a well trained SVM classifier for the testing, and identifying the model of the test vehicle video.

Description

A kind of model recognizing method based on video Fei Sheer vector descriptor

Technical field

The invention belongs to Pattern classification techniques field, particularly a kind of model recognizing method based on video Fei Sheer vector descriptor.

Background technology

The most during the nearly last ten years, the highway traffic infrastructure of China achieves huge construction achievement, and continues developing of high speed.Along with the lasting in-depth of the quick growth of national economy and the urbanization process of China, the vehicle fleet size of China increases swift and violent, brings huge pressure to environment, brings a lot of problem also to urban development and economic growth.Generally, the transport by road cause of China faces following challenge: the population size that (1) constantly expands and the automobile quantity maintained sustained and rapid growth make the increasing pressure of road traffic large; (2) energy consumption of highway transportation is very huge, and energy utilization is insufficient; (3) traffic hazard is occurred frequently, traffic congestion is serious; (4) air environmental pollution is very serious.In order to improve traffic efficiency, improve traffic, present worldwide all uses intelligent transportation system to regulate the traffic in rise.

Core Feature in intelligent transportation system is accurate detection to vehicular traffic and correct vehicle cab recognition.The current research to vehicle detection sorting technique mainly contains two technology schools: automatic vehicle identification and automobile automatic recognition.The former utilizes mobile unit and ground base station equipment to know mutually to carry out, and this technology is mainly used in Fare Collection System, comparatively wide in developed country's usable range, and as the AE-PASS system of the U.S., the ETC system of Japan, Global Satellite GPS locates.The latter is by detecting vehicle parameter inherently, suitable classification and identification algorithm is used under certain vehicle classification standard, on one's own initiative somatotype is carried out to vehicle, this class technology Application comparison is extensive, oneself is applied in real life through there being a lot of ripe system, such technology can identify information of vehicles automatically by modes such as frequency microwave, ruddiness, laser, surface acoustic waves, and the mode of Computer Vision also can be used to identify the information of vehicles such as car plate, vehicle.Comparative maturity technology has Data mining, swashs for infrared detection, ultrasound wave/microwave detection, geomagnetism detecting etc., but this several method respectively has quality, advantage identifies that accurate comparison is high, but shortcoming also clearly, major defect have construction and installation process very complicated, affect normal traffic order, difficult in maintenance, major equipment is fragile, spends larger etc.

In recent years along with the development of Computer Multimedia Technology and image processing techniques, the component that the automobile automatic recognition recognition technology based on video accounts in modern traffic control system is also increasing, and the research effort that various circles of society drop into also gets more and more.Such technology can adapt to the change of dynamic traffic situation, and the traffic flow data a large amount of by Real-time Collection is also transferred to traffic control center, and the data that center is provided by system can make control decision rapidly, and transport solution such as to block up at the problem.Meanwhile, utilize this technology can analyze the information of vehicle flowrate of road, be conducive to general plan and the road construction of network of highways.The advantageous of video detection technology exists: (1) adopts non-contact detecting mode, and installation and maintenance need not break road surface, do not affect pavement life, do not affect traffic; (2) telecommunication flow information in larger scope can be detected, thus reduce number of devices, economize on the use of funds; (3) real time video image of traffic can be provided while gathering telecommunication flow information, be convenient to supervision; (4) for some application, such as traffic census etc., after video image acquisition can being stored, off-line carries out analyzing and processing; (5) when environment changes, or when system moves to elsewhere use, only need simple setting, system can come into operation again.(6) traffic data information and video image be can comprehensively provide, comprehensive, visual inspection to scene are convenient to.

In Images Classification field, often open the local feature of photo for extracting to the very wide method of a kind of application of large-scale image process, the feature of extraction is carried out cluster and coding obtains a high dimension vector, then by it with sorter classification or use nearest neighbor algorithm to mate.The method of wherein encoding has visual word bag model to encode, sparse coding and Fei Sheer vector coding etc.And the performance of Fei Sheer vector coding will be got well compared with other several coded systems, combine the advantage of visual word bag model coding and statistical model, can calculated amount be reduced, reduce memory consumption, and search precision is higher.

On this basis, this patent, in conjunction with the advantage of Fei Sheer vector coding and video detection technology, proposes a kind of model recognizing method based on video Fei Sheer vector descriptor and solves the problems referred to above.The present invention can realize the vehicle cab recognition carrying out vehicle image in video, and recognition accuracy is high, and travelling speed is fast, consumes internal memory few, has higher practicality and robustness.

Summary of the invention

Method described in the present invention is the shortcoming in order to overcome above-mentioned prior art, mainly for the automobile video frequency detected, extract its vehicle characteristics to carry out vehicle segmenting the problem identified, propose a kind of model recognizing method based on video Fei Sheer vector descriptor.Concrete technical scheme is as described below.

For solving the problems of the technologies described above, the present invention by the following technical solutions:

Based on a model recognizing method for video Fei Sheer vector descriptor, comprise the following steps:

Step 1: to the automobile video frequency for training, follows the tracks of the image of each vehicle in video, extracts the SIFT feature of this each two field picture of vehicle;

Step 2: the calculating of Fei Sheer vector coding is carried out to all SIFT feature of the vehicle image extracted, obtains Fei Sheer vector;

Step 3: dimensionality reduction is carried out to the Fei Sheer vector descriptor obtained;

Step 4: carry out binary conversion treatment to the Fei Sheer vector descriptor obtained after dimensionality reduction, obtains the video Fei Sheer vector descriptor of this vehicle;

Step 5: all video Fei Sheer vector descriptors obtained are carried out SVM training, obtains the recognition system that has N number of vehicle classification;

Step 6: to testing vehicle video, the same video Fei Sheer vector descriptor extracting vehicle image in video, tests the model recognition system trained in its steps for importing 5, identifies the vehicle of testing vehicle video.

In technique scheme, to the automobile video frequency for training in described step 1, following the tracks of the image of each vehicle in video, extracting the SIFT feature of this each two field picture of vehicle, comprise following step:

Step 1.1: each vehicle image in track training video;

Step 1.2: data enhancing is carried out to each two field picture of the vehicle traced into, obtains the image data set of often kind of vehicle;

Step 1.3: the SIFT feature extracting the image data set of often kind of vehicle.

In technique scheme, in described step 2, despatch She Er vector coding is carried out to all SIFT feature of the vehicle image extracted and calculates, comprise following step:

Step 2.1: the calculating of Fei Sheer vector coding is carried out to all SIFT feature of vehicle image data set extracted, the Fei Sheer coding vector of a two field picture l that wherein vehicle view data is concentrated is expressed as φ (l);

Step 2.2: the process of symbol square root is carried out to Fei Sheer coding vector φ (l) of every two field picture l:

$sign\left(\mathrm{\φ}\right(l\left)\right)/\sqrt{\left|\mathrm{\φ}\left(l\right)\right|},$

Again L2 normalization is carried out to the vector after the process of symbol square root.

In technique scheme, described step 3 carries out PCA dimensionality reduction to the Fei Sheer vector descriptor obtained of encoding, and comprises study projection matrix W, by Fei Sheer vector descriptor from R ^ddimension is down to R ^mdimension.

In technique scheme, the Fei Sheer vector descriptor after described step 4 pair dimensionality reduction carries out binary conversion treatment, obtains the video Fei Sheer vector descriptor of this vehicle, comprises following step:

Step 4.1: the Fei Sheer vector descriptor calculated in step 3, forms a matrix: U ∈ R ^{q × m};

Step 4.2: by sign function, sign (a)=1, iffa > 0 carries out binaryzation, obtains a binaryzation vector:

β＝sign(Uψ-w)

Wherein ψ is the Fei Sheer vector descriptor after dimensionality reduction, and w is selected threshold value, as U ψ > w, and β=1, otherwise β=0;

Step 4.3: binaryzation coded descriptor β={ 0, the 1} obtaining a q position ^q, the also video Fei Sheer vector descriptor of i.e. this vehicle.

In technique scheme, all descriptors obtained are carried out SVM training by described step 5, obtain the recognition system that has N number of vehicle classification, comprise and use onevsrestSVM to train a multi classifier, obtain the recognition system with N class classification.

In sum, owing to have employed technique scheme, the invention has the beneficial effects as follows:

The invention discloses a kind of model recognizing method directly to automobile video frequency process, vehicle to be identified can be detected fast, take Fei Sheer vector coding to represent the vehicle image in video, calculated amount can be reduced, reduce memory consumption, and search precision is higher, improves vehicle cab recognition rate.

Accompanying drawing explanation

Examples of the present invention will be described by way of reference to the accompanying drawings, wherein:

Fig. 1 is that video Fei Sheer vector algorithm realizes schematic diagram.

Embodiment

By describing technology contents of the present invention, structural attitude in detail, realized object and effect, accompanying drawing is coordinated to be explained in detail below in conjunction with embodiment.

The present invention proposes a kind of model recognizing method based on video Fei Sheer vector descriptor, vehicle vehicle cab recognition obtains good effect.Whole algorithm realization schematic diagram as shown in Figure 1, comprises step:

Step 1: to the automobile video frequency for training, follows the tracks of the image of each vehicle in video, extracts the SIFT feature of this each two field picture of vehicle;

Mainly comprise following step:

Step 1.1: each vehicle image in track training video, because the present invention is mainly in vehicle cab recognition, what tracking module was taked is the image that basic particle filter technology follows the tracks of each vehicle in video, does not go deep at this;

Step 1.2: data enhancing is carried out to each two field picture of the vehicle traced into, obtains the image data set of often kind of vehicle;

Step 1.3: the SIFT feature extracting the image data set of often kind of vehicle.

Current tracking technique has a variety of, and the present invention takes basic particle filter technology to follow the tracks of the image of each vehicle in video, extracts the SIFT feature of this each two field picture of vehicle; Extract SIFT feature and comprise following step:

Step 1.4: metric space extremum extracting: for the gaussian pyramid image built, utilize difference function to detect candidate's extreme point on all yardsticks.

Step 1.5: key point is located: accurate location and the yardstick of being located extreme point by Function Fitting, filter out the marginal point in the set of candidate's extreme point simultaneously.

Step 1.6: direction is determined: in conjunction with the gradient information of key point neighborhood territory pixel, gives a principal direction to each key point.

Step 1.7: Feature Descriptor generates: the gradient magnitude of key point neighborhood territory pixel and direction are added up, obtains the feature interpretation of key point.

Step 2: the calculating of Fei Sheer vector coding is carried out to all SIFT feature of the vehicle image extracted;

The essence of Fei Sheer vector coding utilizes the gradient vector of the likelihood function of image to represent piece image.Piece image can be made up of a lot of local feature, is independently between each dimension of each feature, can simulate these local features with probability model, and like this, the probability distribution of piece image can be expressed as the product of the probability distribution in each characteristic dimension.

The probability model that this patent uses is gauss hybrid models, according to the difference of Gaussian probability-density function parameter, each Gauss model regards a classification as, inputs a sample, calculate the value of its probability density function, carry out judgement sample by threshold value and whether belong to this Gauss model.For the Fei Sheer vector coding of mixed Gauss model, being exactly the Gaussian distribution representing image, respectively local derviation being asked to its weight, expectation and covariance, representing that image comprises following step by the result obtained:

1.X={x _t, t=1 ... T} is expressed as the description vectors of an image, x in this patent _trepresent each SIFT descriptor, λ={ ω _i, μ _i, Σ _i, i=1 ... N} is parameter set, is expressed as the weight of the Gaussian function of analog image SIFT descriptor, expectation and covariance matrix;

2. be independent identically distributed between each SIFT feature, then have the probability density function of gauss hybrid models to be:

$P\left(X\right|\mathrm{\λ})=\underset{t=1}{\overset{T}{\Π}}p\left(x_t\right|\mathrm{\λ})$

Wherein p (x _t| λ) represent gauss hybrid models, have:

$p\left(x_t\right|\mathrm{\λ})=\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\ω}}_i{p}_i\left(x_t\right|\mathrm{\λ})$

ω _irepresent the weights of each single Gaussian function, p _i(x _t| λ) represent each single Gauss model, have:

$p_i\left(x_t\right|\mathrm{\λ})=\frac{\exp \{-\frac{1}{2}{(x_t-{\mathrm{\μ}}_i)}^T{\mathrm{\Σ}}_i^{-1}(x_t-{\mathrm{\μ}}_i)\}}{{\left(2\mathrm{\π}\right)}^{D/2}|{\mathrm{\Σ}}_i{|}^{1/2}}$

Wherein D represents the dimension of SIFT feature point, and the log-likelihood function of gauss hybrid models is:

$L\left(X\right|\mathrm{\λ})=\underset{t=1}{\overset{T}{\Σ}}log\left\{\underset{i=1}{\overset{N}{\Σ}}{w}_i{P}_i\left(x_t\right|\mathrm{\λ})\right\}$

3. the unique point x of image _tthe probability generated by i-th Gaussian distribution is:

${\mathrm{\γ}}_t\left(i\right)=\frac{{\mathrm{\ω}}_i{p}_i\left(x_t\right|\mathrm{\λ})}{\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\ω}}_j{p}_j\left(x_t\right|\mathrm{\λ})}$

4. the Gaussian mixtures of his-and-hers watches diagram picture, respectively local derviation is asked to its weight, expectation and covariance:

$\frac{\∂L\left(X\right|\mathrm{\λ})}{\∂{\mathrm{\ω}}_i}=\underset{t=1}{\overset{T}{\Σ}}\[\frac{{\mathrm{\γ}}_t\left(i\right)}{{\mathrm{\ω}}_i}-\frac{{\mathrm{\γ}}_t\left(1\right)}{{\mathrm{\ω}}_1}\]---(i\≥2)$

$\frac{\∂L\left(X\right|\mathrm{\λ})}{\∂{\mathrm{\μ}}_i^d}=\underset{t=1}{\overset{T}{\Σ}}{\mathrm{\γ}}_t\left(i\right)\[\frac{x_t^d-{\mathrm{\μ}}_i^d}{{\left({\mathrm{\σ}}_i^d\right)}^2}\]$

$\frac{\∂L\left(X\right|\mathrm{\λ})}{\∂{\mathrm{\σ}}_i^d}=\underset{t=1}{\overset{T}{\Σ}}{\mathrm{\γ}}_t\left(i\right)\[\frac{{(x_t^d-{\mathrm{\μ}}_i^d)}^2}{{\left({\mathrm{\σ}}_i^d\right)}^3}-\frac{1}{{\mathrm{\σ}}_i^d}\]$

Wherein d is the dimension of feature, and σ is the standard deviation of Gaussian distribution;

5. will the vector after local derviation be asked to carry out standardization, obtains final Fei Sheer coding vector:

$f_{{\mathrm{\ω}}_i}^{-1/2}\∂L\left(X\right|\mathrm{\λ})/\∂{\mathrm{\ω}}_i$

$f_{{\mathrm{\μ}}_i^d}^{-1/2}\∂L\left(X\right|\mathrm{\λ})/\∂{\mathrm{\μ}}_i^d$

$f_{{\mathrm{\σ}}_i^d}^{-1/2}\∂L\left(X\right|\mathrm{\λ})/\∂{\mathrm{\σ}}_i^d$

Wherein $f_{{\mathrm{\ω}}_i}=T(\frac{1}{{\mathrm{\ω}}_i}+\frac{1}{{\mathrm{\ω}}_1}),f_{{\mathrm{\μ}}_i^d}=\frac{{\mathrm{T\ω}}_i}{{\left({\mathrm{\σ}}_i^d\right)}^2},f_{{\mathrm{\σ}}_i^d}=\frac{2{\mathrm{T\ω}}_i}{{\left({\mathrm{\σ}}_i^d\right)}^2}.$

Step 2.1: the calculating of Fei Sheer vector coding is carried out to all SIFT feature of vehicle image data set extracted, the Fei Sheer coding vector of a two field picture l that wherein vehicle view data is concentrated is expressed as φ (l);

Step 2.2: the process of symbol square root is carried out to Fei Sheer coding vector φ (l) of every two field picture l:

$sign\left(\mathrm{\φ}\right(l\left)\right)/\sqrt{\left|\mathrm{\φ}\left(l\right)\right|},$

Again L2 normalization is carried out to the vector after the process of symbol square root.

Step 3: dimensionality reduction is carried out to the Fei Sheer vector descriptor obtained;

The dimension of the Fei Sheer vector descriptor obtained through step 2 is still very high, and the complexity of calculating is too large, by PCA dimensionality reduction, learns a projection matrix W, Fei Sheer vector descriptor is down to R ^mdimension.

Step 4: carry out binary conversion treatment to the descriptor after dimensionality reduction, obtains the video Fei Sheer vector descriptor of this vehicle;

After step 3 dimensionality reduction, in order to further reduce the consumption of internal memory, also need to be optimized, by the real-valued vectors descriptor ψ ∈ R after dimensionality reduction ^mbe mapped to binaryzation coding β={ 0 a, 1} ^q, comprise following step:

Step 4.1: the Fei Sheer vector descriptor calculated in step 3, forms a matrix U ∈ R ^{q × m};

Step 4.2: by sign function, sign (a)=1, iffa > 0 carries out binaryzation, obtains a binaryzation vector:

β＝sign(Uψ-w)

Wherein ψ is the Fei Sheer vector descriptor after dimensionality reduction, and w is selected threshold value, as U ψ > w, and β=1, otherwise β=0;

Step 4.3: binaryzation coded descriptor β={ 0, the 1} obtaining a q position ^q, the also video Fei Sheer vector descriptor of i.e. this vehicle.

Step 5: all descriptors obtained are carried out SVM training, obtains the recognition system that has N number of vehicle classification;

Obtain the video Fei Sheer vector descriptor of training vehicle by step 1 to 4, carried out the training of onevsrest Linear SVM, obtain the recognition system that has N number of vehicle classification.

Step 6: the video Fei Sheer vector descriptor extracting testing vehicle video, is imported the SVM classifier trained and test, identify the vehicle of testing vehicle video.

For the video of testing vehicle, same its video of extraction Fei Sheer vector descriptor, the recognition system trained before being imported identifies, comprises following step:

Step 6.1: by step 1 to step 4, extracts the video Fei Sheer vector descriptor of testing vehicle video;

Step 6.2: the video Fei Sheer vector descriptor of extraction is imported the SVM classifier trained and tests, identify the vehicle of testing vehicle video.

Claims (6)

1., based on a model recognizing method for video Fei Sheer vector descriptor, comprise the following steps:

Step 1: to the automobile video frequency for training, follows the tracks of the image of each vehicle in video, extracts the SIFT feature of this each two field picture of vehicle;

Step 2: the calculating of Fei Sheer vector coding is carried out to all SIFT feature of the vehicle image extracted, obtains Fei Sheer vector;

Step 3: PCA dimensionality reduction is carried out to the Fei Sheer vector descriptor obtained;

Step 4: carry out binary conversion treatment to the Fei Sheer vector descriptor obtained after dimensionality reduction, obtains the video Fei Sheer vector descriptor of this vehicle;

Step 5: all video Fei Sheer vector descriptors obtained are carried out SVM training, obtains the recognition system that has N number of vehicle classification;

Step 6: to testing vehicle video, the same video Fei Sheer vector descriptor extracting vehicle image in video, tests the model recognition system trained in its steps for importing 5, identifies the vehicle of testing vehicle video.

2. according to claim 1 based on the model recognizing method of video Fei Sheer vector descriptor, it is characterized in that, to the automobile video frequency for training in described step 1, following the tracks of the image of each vehicle in video, extract the SIFT feature of this each two field picture of vehicle, comprise following step:

Step 1.1: each vehicle image in track training video;

Step 1.2: data enhancing is carried out to each two field picture of the vehicle traced into, obtains the image data set of often kind of vehicle;

Step 1.3: the SIFT feature extracting the image data set of often kind of vehicle.

3. according to claim 1 based on the model recognizing method of video Fei Sheer vector descriptor, it is characterized in that, in described step 2, the calculating of Fei Sheer vector coding carried out to all SIFT feature of the vehicle image extracted, comprise following step:

Step 2.1: the calculating of Fei Sheer vector coding is carried out to all SIFT feature of vehicle image data set extracted, the Fei Sheer coding vector of a two field picture l that wherein vehicle view data is concentrated is expressed as φ (l);

Step 2.2: the process of symbol square root is carried out to Fei Sheer coding vector φ (l) of every two field picture l:

$sign\left(\mathrm{\φ}\right(l\left)\right)/\sqrt{\left|\mathrm{\φ}\left(l\right)\right|},$

Again L2 normalization is carried out to the vector after the process of symbol square root.

4. according to claim 1 based on the model recognizing method of video Fei Sheer vector descriptor, it is characterized in that, described step 3 carries out PCA dimensionality reduction to the Fei Sheer vector descriptor obtained of encoding, and comprises study projection matrix W, by Fei Sheer vector descriptor from R ^ddimension is down to R ^mdimension.

5. according to claim 1 based on the model recognizing method of video Fei Sheer vector descriptor, it is characterized in that, Fei Sheer vector descriptor after described step 4 pair dimensionality reduction carries out binary conversion treatment, obtains the video Fei Sheer vector descriptor of this vehicle, comprises following step:

Step 4.1: the Fei Sheer vector descriptor calculated in step 3, forms a matrix: U ∈ R ^{q × m};

Step 4.2: by sign function, sign (a)=1, iffa > 0 carries out binaryzation, obtains a binaryzation vector:

β＝sign(Uψ-w)

Wherein ψ is the Fei Sheer vector descriptor after dimensionality reduction, and w is selected threshold value, as U ψ > w, and β=1, otherwise β=0;

Step 4.3: binaryzation coded descriptor β={ 0, the 1} obtaining a q position ^q, the also video Fei Sheer vector descriptor of i.e. this vehicle.

6. according to claim 1 based on the model recognizing method of video Fei Sheer vector descriptor, it is characterized in that, the all descriptors obtained are carried out SVM training by described step 5, obtain the recognition system that has N number of vehicle classification, comprise and use onevsrestSVM to train a multi classifier, obtain the recognition system with N class classification.