CN100520853C - Vehicle type classification method based on single frequency continuous-wave radar - Google Patents

Abstract

The vehicle classification method includes: the time-domain radar signal of single-frequency continuous wave radar is used as input to conduct time-frequency analysis to obtain a radar echo Doppler frequency spectrum chart changing along time; by Hafford counterchange the Doppler chart is mapped into a scattered central position parameter image; through characteristic picking-up, Karhunen-Loeve(K-L) screening and compressing, a characteristic sample is obtained, which is classified using Fisher rule to obtain the vehicle-type classification. For multi-type condition, multiple two-type classifiers are used to conduct respectively pairwise classification, according to the result of the classifiers said above the classification result is obtained by voting. Advantage: simple and reliable structure and low cost, conducting classification synchronously with detecting and speed-measuring of vehicles.

Description

Vehicle type classification method based on single-frequency continuous wave radar

Technical field

Single-frequency continuous wave radar adopts the working method of continuous launched microwave signal, uses zero intermediate frequency reciver, and local oscillation signal and received signal direct conversion obtain Doppler signal.Driving vehicle detects, carries out vehicle classification when testing the speed to the present invention relates to satisfy the need upward with single-frequency continuous wave radar, belongs to the intelligent traffic vehicle detection range.

Background technology

The present invention is directed to current vehicles and detect radar and can only find range and test the speed and can't carry out the problem of vehicle exhaustive division vehicle, the vehicle doppler information that has proposed to utilize single-frequency continuous wave radar cheaply to provide carries out the method for vehicle classification.

Following article and patent documentation have covered the main background technology in this field substantially.In order to explain out the evolution of technology, allow their time sequencings arrange, and introduce the main contribution and the shortcoming of document one by one.

1.Roe?H，Hobson?G.，Improved?Discrimination?of?Microwave?Vehicle?Profiles，IEEEMTT-S?International?Microwave?Symposium?Digest?1992，1992，2：717-720

The sorting technique that distance and doppler information combine, adopt Continuous Wave with frequency modulation (FMCW) technology, the author adopts that a bandwidth is big, the narrower frequency modulated continuous wave radar of wave beam is installed in the road top, angle lapping at 45 is tiltedly observed a specific cell territory on the road, measure the distance and the speed of moving target simultaneously, on the x-y plane, recover the profile of vehicle according to time, speed and range information, carry out the classification of vehicle on this basis.Shortcoming is to the having relatively high expectations of radar equipment, and has hindered this The Application of Technology and popularization.

2.Martinez?M，Casar?C，Miguel?VG，A?Neural?Network?Approach?to?Doppler-basedTarget?Classification，Proceeding?of?the?International?Conference?on?Rader?92，1992，450-453

Based on the sorting technique of Doppler's feature, utilize neural network method to distinguish aircraft/vehicle (Aircraft/Vehicle), helicopter (Helicopter), people (Persons) and clutter (Clutter).Being applied to the far field radar, is not special classification at the ordinary road vehicle.

3.Jahangir?M，Ponting?K，JW.0.，A?robust?Doppler?Classification?Technique?basedon?Hidden?Markov?Models，Proceedings?of?the?Internation?Conference?on?Radar?2002，2002，162-166

Based on the sorting technique of Doppler's feature, utilize hidden Markov model (HMM) differentiation personnel (Personnel), creeper truck (Tracked Vehicle) and wheeled vehicle (Wheeled Vehicle) tertiary target.Being applied to the far field radar, is not special classification at the ordinary road vehicle.

4.Urazghildiiev?I，Ragnarsson?R，Wallin?K，et?al，A?Vehicle?ClassificationSystem?Based?on?Microwave?Radar?Measurement?of?Height?Profi?le，Proceedings?of?theInternational?Conference?on?Radar?2002，2002，409-413

Sorting technique based on height profile., through out-of-date vehicle body is scanned at vehicle, the height that obtains the vehicle body part can obtain height profile very clearly over time.This scheme adopts the wideband pulse radar usually, and structure is comparatively complicated, and cost is higher, has brought certain degree of difficulty for the popularization of this technology.

5.Xuan?Y，Meng?H，Wang?X，et?al，A?High-Range-Resolution?Microwave?Radar?Systemfor?Traffic?Flow?Rate?Measurement，Proceedings?of?the?8 ^th?IEEE?Conference?onIntelligent?Transportation?Systems，2005，880-885

Based on the sorting technique of height profile, use a bandwidth big, the radar that wave beam is narrow, be installed in carriage way directly over vertically downward observation.Because it has very high range resolution (corresponding bandwidth) and spatial resolution (corresponding beam angle), thereby at vehicle through out-of-date, can scan vehicle body, the height that obtains the vehicle body part variation of the lengthwise position of vehicle body (just along with) in time has very high range resolution, can obtain height profile very clearly.The shortcoming radar is perpendicular to floor mounted, and the wave beam of radar is vertical with the direction of motion of vehicle target, thereby can not utilize the speed of doppler information measuring vehicle; Secondly, obtain vehicle ' s contour information accurately from above-mentioned measurement result (variation diagram highly in time), must know the speed of a motor vehicle, the uncertainty of the speed of a motor vehicle can be brought certain bluring to vehicle identification; At last, this scheme adopts the wideband pulse radar usually, and structure is comparatively complicated, and cost is higher, has brought certain degree of difficulty for the popularization of this technology.

6.Bilik?I，Tabrikian?J，Cohen?A，Target?Classification?using?Gaussian?Mixturemodel?for?Ground?Surveillance?Doppler?Radar，Proceedings?of?2005?IEEE?InternationalRadar?Conference，2005，910-915

Based on the sorting technique of Doppler's feature, the generation of Doppler's feature fundamentally has two class reasons: a class Doppler feature is caused by the internal motion of target, is characterized in that Doppler signal has a plurality of frequency components at synchronization.Another kind of reason is that the characteristics of motion of target has time dependent characteristics, for example some target is not to do desirable uniform motion, but speed in time or the locus do periodic fluctuation, such characteristics show and have just formed the characteristics that Doppler frequency fluctuates in time on the Doppler signal.Above-mentioned two class reasons combine, and have caused different Doppler's features of different target type, can classify to them in view of the above.It or not special classification at the ordinary road vehicle.

7.Kjellgren?J，Gadd?S，Jonsson?N，et?al，Analysis?of?Doppler?Measurements?ofGround?Vehicles，Proceedings?of?2005?IEEE?International?Radar?Conference，2005，284-289

Based on the sorting technique of Doppler's feature, analyzed Doppler's feature of ground endless-track vehicle by emulation and actual radar data; Utilize gauss hybrid models (GMM) realization classification on a surface target, classification comprises pedestrian, wheeled vehicle, creeper truck, animal and clutter etc.Being applied to the far field radar, is not special classification at the ordinary road vehicle.

Summary of the invention

By to the existing traffic flow detection and the summary of sorting technique and more as can be seen based on microwave radar,, existing sorting technique realizes classification to vehicle target thereby all being the direct or indirect estimation vehicle ' s contour information of information such as distance, speed or Doppler's feature of utilizing radar to record.Use pulsed radar or frequency conversion continuous wave radar cost higher, be unfavorable for the widespread use popularization, and the microwave current radar application still is in the research starting stage in the traffic flow sorting technique.

Find through summing up experimental data, Doppler's spectral line pattern of vehicle echoed signal is by the decision of the space distribution of target vehicle scattering center, and the distribution of scattering center is determined by the vehicle body shape, therefore, if can calculate the scattering center of vehicle from Doppler's power spectrum of target vehicle echoed signal distributes, the i.e. shape of estimating target vehicle and it is carried out type identification in view of the above, visible Doppler's feature can be directly used in the classification of vehicle target.

The present invention is based on the vehicle type classification method of single-frequency continuous wave radar, with the time domain radar signal of single-frequency continuous wave radar as input, carry out time frequency analysis and obtain time dependent radar return spectrogram, is vehicle scattering center location parameter image by hough transform with the Doppler frequency spectrum image mapped, obtain feature samples through feature extraction and compression, utilize the Fisher criterion to carry out sample classification, obtain the vehicle classification of vehicle, performing step is as follows:

Step 1 is carried out Doppler frequency that time frequency analysis obtains echoed signal modified-image in time with the time domain radar signal.Adopt short time discrete Fourier transform STFT:

$\mathrm{STFT}\{x\[\]\}\≡X(m,\mathrm{\ω})=\underset{n=-\∞}{\overset{+\∞}{\mathrm{\Σ}}}x\left[n\right]w[n-m]e^{-\mathrm{j\ωn}}$

X[n wherein] be discrete-time signal, w[n] be window function, m is the sliding position of window function, ω is an angular frequency.The result of STFT is a distribution on the Time And Frequency two dimensional surface, get STFT the result mould square, expression input signal x[n] distribute power on the Time And Frequency plane, represent with spectrogram.

The Doppler frequency of the echoed signal that the vehicle scattering center produces is over time:

$f_d\left(t\right)=\frac{2\mathrm{vf}}{c}\frac{x_0+x_v-\mathrm{vt}}{\sqrt{{(x_0+x_v-\mathrm{vt})}^2+{(h_0-h_v)}^2}}$

Wherein, x _vBe vehicle scattering center horizontal ordinate, h _vBe vehicle scattering center ordinate, x ₀Be vehicle admission position, h ₀Be the radar setting height(from bottom), f is the radar frequency of operation, and c is the light velocity, f _dBe the Doppler frequency of vehicle scattering center reflection echo, v is the vehicle relative velocity.

The distribution of car scattering center experience and corresponding Doppler's spectral line are as shown in Figure 1.Fig. 1 is vehicle scattering center and corresponding Doppler curve synoptic diagram.The left figure of Fig. 1 marks the vehicle scattering center, is parameter x _v-h _vThe parameter space that constitutes.Right figure curve is the Doppler frequency spectrum curve of the vehicle scattering center reflection echo of correspondence, is Doppler frequency spectrum figure t-f _dThe image space that constitutes.

Step 2 utilizes hough transform with Doppler frequency spectrum figure t-f _dImage space is converted to parameter x _v-h _vThe space obtains parameter x _v-h _vThe distribution of vehicle scattering strength in the space, the curve in the parameter space is:

$h_v\left(x_v\right)=h_0-(x_v+x_0-\mathrm{vt})\sqrt{{\left(\frac{2\mathrm{vf}}{f_{d}c}\right)}^2-1}$

Straight line in the corresponding parameter space of point on the spectrogram curve, the cluster straight line in the corresponding parameter space of a curve on the spectrogram, the intersection point (x of this cluster straight line _v, h _v) be exactly the parameter of Doppler's spectral line, the corresponding scattering center coordinate of vehicle just.Fig. 2 is the hough transform synoptic diagram.Each bar Doppler curve process hough transform of left figure becomes the many bunches of straight lines of right figure among Fig. 2, and the intersection point of every bunch of straight line is exactly the vehicle scattering center.

In the process that realizes hough transform, obtain x by the method for shining upon _v-h _vDistribution R (the x of scattering strength in the plane _v, h _v), this distribution is obtained by integration:

$R(x_v,h_v)={\∫}_{t_0}^{t_1}{X}_P(t,f_d)\mathrm{dt}={\∫}_{t_1}^{t_1}{X}_P(t,\frac{2\mathrm{vf}}{c}\frac{x_0+x_v-\mathrm{vt}}{\sqrt{{(x_0+x_v-\mathrm{vt})}^2+{(h_0-h_v)}^2}})\mathrm{dt}$

Wherein, X _PBe the power spectrum of radar return:

$X_p[m,k]={\left|\mathrm{FFT}\left\{x[m,i]w\left[i\right]\right\}\right|}^2={\left|\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}x[m,i]w\left[i\right]W_N^{\mathrm{nk}}\right|}^2$

X[m wherein, i] be radar return sampled signal x[n] by fixed step size intercepting, w[i] be window function.

In the reality, because multiple factor affecting, the parameter space image that the result of hough transform obtains is respectively organized straight line and is met at a bit unlike ideal situation, forms fuzzy belt-like zone but intersect.But the distribution shape of dash area can reflect that still the vehicle scattering center gets the shape of characteristic distributions and vehicle body among the result.

Step 3, feature extraction: because factors such as radar setting angles, vehicle scatter intensity distribution figure striped can be angled, at first scatter intensity distribution figure striped rotation respective angles is adjusted into vertical direction, and the horizontal direction intensity segment information that can extract striped like this is as characteristic of division.Fig. 3 adjusts synoptic diagram for scatter intensity distribution figure.

The maximal value of getting each perpendicular row then obtains the one-dimensional characteristic vector space that can represent scatter intensity distribution.With R (x _v, h _v) discretize obtains R[x _v, h _v], get R[x _v, h _v] in the maximal value of each row, carry out normalization then, intensity distribution image is mapped as an one-dimensional vector r[x]:

$r\left[x\right]=\frac{\underset{h}{\max }R[x,h]}{\underset{x,h}{\max }R[x,h]}$

Result as shown in Figure 4.Fig. 4 is a vehicle scatter intensity distribution figure result.

Step 4, feature screening and compression: because the characteristic vector space length that as above obtains is bigger, wherein there is more invalid data can cause problems such as the excessive and numerical value instability of calculated amount, further screens and compress, find out a spot of the most effective feature of classification to this proper vector r.

It is generally acknowledged in the sample set of no label, have also big than degree of separation between its class of characteristic quantity of big variance, therefore adopt Karhunen-Loeve (K-L) to launch, with eigenvector projection in new feature space by variance size ordering, also just be equal to ordering, therefrom choose the top n proper vector that can reflect feature and keep better classifying quality according to the class discrimination ability.

The Karhunen-Loeve transformation process is as follows:

At first computational data intensity distributions proper vector set the covariance matrix ∑ of r}:

∑=cov (r)=E[(r-μ) (r-μ) ^T], wherein μ is the population mean vector of feature samples collection.

Adopt the latent vector of the method compute matrix ∑ of svd (SVD) then:

$\mathrm{\Σ}={\mathrm{U\Λ}}^{\frac{1}{2}}{V}^T,$ Wherein U, V are orthogonal matrix, are made up of the latent vector of ∑ matrix.Λ is a diagonal matrix, and element is the eigenvalue of the ∑ matrix arranged from big to small on the diagonal line.

Utilize latent vector U as orthogonal basis, sample set can be projected in the new feature space by the ordering of variance size:

X＝U ^Tr。

In the SVD process, eigenvalue is arranged according to order from big to small, therefore in new feature space, the variance of each latitude sample set is successively decreased, promptly in vectorial X, the feature variance that comes the front is big, contain much information, and comes that the latter feature variance is little, quantity of information is little.For the latitude that reduces characteristic quantity, reduce calculated amount, can directly from X, choose the top n feature and give up remaining, use after the Karhunen-Loeve transformation characteristic quantity information minimum of way loss like this.

Step 5 is compared characteristic quantity by criterion, finishes classification.After the compression of having finished target sample Feature Extraction and characteristic quantity space dimensionality, can adopt simple Fisher criterion to carry out the linear classification of two class vehicles, can realize the polymorphic type classification with a plurality of such sorting techniques then.

Utilize to obtain normalized optimal weight vector under the Fisher criterion, utilize optimal weight vector that sample is projected to the one-dimensional space and compares the affiliated type of judgement sample with average optimal threshold power.

The method of two class line styles classification is as follows:

For the sample vector X of N dimensional feature space, find weight vector w and threshold value power w ₀, make discriminant function:

G (X)=w ^TX+w ₀Satisfy decision rule:

L wherein ₁, L ₂Be respectively the sample set of two classes.

The process of determining weight vector w is as follows:

Definition m _iBe sample average vector, S _iBe sample within class scatter matrix, S _LBe total within class scatter matrix:

$m_i=\frac{1}{N}\underset{X\∈L_i}{\mathrm{\Σ}}X,$ i＝1，2

$S_i=\underset{X\∈L_i}{\mathrm{\Σ}}(X-m_i){(X-m_i)}^T,$ i＝1，2

S _L＝S ₁+S ₂

Then normalized optimal weight vector w is under the Fisher criterion:

$w=\frac{S_L^{-1}(m_1-m_2)}{\left|\right|S_L^{-1}(m_1-m_2)\left|\right|}$

Obtain sample to be projected to one-dimensional space Y after the optimal weight vector w, and utilize average method to calculate optimal threshold power Y ₀

Y _n＝w ^TX _n，n＝1，2，…，N _i

$m_i=\frac{1}{N_i}\underset{X\∈L_i}{\mathrm{\Σ}}Y,$ i＝1，2

${\mathrm{<figure-callout\; id="0"\; label="Y"\; filenames="C200710175796E00111.png,C200710175796E00121.png"\; state="\{\{state\}\}">Y</figure-callout>}}_0=\frac{{\tilde{m}}_1+{\tilde{m}}_2}{2}$

So far, for arbitrary sample X, can use optimal weight vector w to calculate its projection Y, then to Y and Y ₀Size compare, can adjudicate the type of X.

Re-use a plurality of two classification of type devices for sample polymorphic type situation and carry out pairwise classification respectively, the result according to sorter votes then, adopts sample subpoint in different sorters to arrive the distance D of boundary threshold point _i=| Y _i-Y _0i|, select maximum D _i, as final classification results.

Single-frequency continuous wave radar has advantage simple and reliable for structure, with low cost, has been applied to the detection of vehicle target and tests the speed.The present invention utilizes single-frequency continuous wave radar to realize in vehicle detection, carries out vehicle classification when testing the speed, and the average classification accuracy rate that in the practical application vehicle is divided three classes can reach 94.8%.

Description of drawings

Fig. 1 is vehicle scattering center and corresponding Doppler curve synoptic diagram.

Fig. 2 is the hough transform synoptic diagram.

Fig. 3 adjusts synoptic diagram for scatter intensity distribution figure.

Fig. 4 is a vehicle scatter intensity distribution figure result.

Fig. 5 process flow diagram of the present invention.

Fig. 6 is short time discrete Fourier transform (STFT) result.

Fig. 7 is the hough transform result.

Fig. 8 is the feature extraction result.

Embodiment

Below in conjunction with embodiment the present invention is described.

Fig. 5 is a process flow diagram of the present invention.As shown in Figure 5, the software of radar traffic current sensor is made up of a plurality of functional modules such as range observation, vehicle detection, velocity estimation, vehicle classification, scene perception, State Control and input and output.Among Fig. 5, the Signal Pretreatment stage comprises above-mentioned steps one; The prime of detection algorithm stage sorting algorithm provides information such as vehicle turnover field time, travel speed; The sorting algorithm stage comprises that step 2 is to step 5;

Technical at the existed system model machine, we have carried out experiment and data acquisition on the spot on the Chengfu Road, Haidian District, Beijing City, and 45 ° of radar depression angles, installation site far from the road surface 6 meters have obtained 164 vehicle samples after data processing.

Fig. 6 is short time discrete Fourier transform (STFT) result.A figure is the ripple signal among Fig. 6, and B figure is the echo Doppler frequency spectrum that obtains after handling through short time discrete Fourier transform, and C schemes the video pictures of echo corresponding vehicle for this reason.Fig. 6 is short time discrete Fourier transform (STFT) result.Echo Doppler frequency spectrum figure is carried out hough transform obtain vehicle scatter intensity distribution figure, Fig. 7 is the hough transform result, and left figure is the Doppler frequency spectrum figure of different vehicle among Fig. 7, and right figure breathes out corresponding vehicle scatter intensity distribution figure after the not conversion.

Vehicle scatter intensity distribution figure is carried out characteristic image after the screening of angular setting and feature obtains dimensionality reduction, proper vector image such as Fig. 8 of corresponding vehicle scattering center.Fig. 8 is the feature extraction result.

So far obtain 520 proper vector r, further compress screening and compression, choose preceding 14 after launching with K-L and kept enough features and made the compromise proper vector of the less effect of calculated amount as grouped data.

According to vehicle characteristics and data characteristics, with sample be divided into 3 classes: A small-sized (car, SUV); B medium-sized (minibus); C large-scale (motor bus, lorry).

Use the Fisher linear decision rule to realize three independently two class linear classifiers, be respectively applied for and differentiate A/B, A/C, B/C, the result votes and produces final classification results.

164 vehicle samples are split as training set and test set at random, sorter is carried out training and testing, the statistical classification result has repeatedly calculated 94.8% average classification accuracy rate.Concrete classification results is as shown in the table:

	Sum	Be divided into 1 class	Be divided into 2 classes	Be divided into 3 classes	Accuracy
						True
1 class	135	128.95	4.53	1.52	95.5％
						True 2 classes	17	1.35	15.56	0.09	91.5％
True 3 classes	12	0.70	0.31	10.99	91.6％
						Add up to	164	-	-	-	94.8％

Experimental result has proved that the microwave traffic flow detection of our design and categorizing system and corresponding signal process algorithm are effective and practical.

Claims (2)

1. Based on the vehicle type classification method of single-frequency continuous wave radar, it is characterized in that 1, this method step is as follows:

   Step 1 is carried out the Doppler frequency spectrum figure t-f that time frequency analysis obtains echoed signal with the time domain radar signal _dImage space:

   Adopt short time discrete Fourier transform STFT:

   $\mathrm{STFT}\left\{x\right[\left]\right\}\&equiv;X(m,\mathrm{\&omega;})=\underset{n=-\&infin;}{\overset{+\&infin;}{\mathrm{\&Sigma;}}}x\left[n\right]w[n-m]e^{-\mathrm{j\&omega;n}}$

   X[n wherein] be discrete-time signal, w[n] be window function, m is the sliding position of window function, ω is an angular frequency; The Doppler frequency spectrum figure t-f of the echoed signal that the vehicle scattering center produces _dImage space is:

   $f_d\left(t\right)=\frac{2\mathrm{vf}}{c}\frac{x_0+x_v-\mathrm{vt}}{\sqrt{{(x_0+x_v-\mathrm{vt})}^2+{(h_0-h_v)}^2}}$

   Wherein, x _vBe vehicle scattering center horizontal ordinate, h _vBe vehicle scattering center ordinate, x ₀Be vehicle admission position, h ₀Be the radar setting height(from bottom), f is the radar frequency of operation, and c is the light velocity, f _dBe the Doppler frequency of vehicle scattering center reflection echo, v is the vehicle relative velocity;

   Step 2 utilizes hough transform with Doppler frequency spectrum figure t-f _dImage space is converted to parameter x _v-h _vThe space obtains parameter x _v-h _vThe distribution of vehicle scattering strength in the space, the curve in the parameter space is:

   $h_v\left(x_v\right)=h_0-(x_v+x_0-\mathrm{vt})\sqrt{{\left(\frac{2\mathrm{vf}}{f_{d}c}\right)}^2-1};$

   In the process that realizes hough transform, obtain x by the method for shining upon _v-h _vDistribution R (the x of scattering strength in the plane _v, h _v), this distribution is obtained by integration:

   $R(x_v,h_v)={\&Integral;}_{t_0}^{t_1}{X}_P(t,f_d)\mathrm{dt}={\&Integral;}_{t_1}^{t_1}{X}_P(t,\frac{2\mathrm{vf}}{c}\frac{x_0+x_v-\mathrm{vt}}{\sqrt{{(x_0+x_v-\mathrm{vt})}^2+{(h_0-h_v)}^2}})\mathrm{dt},$

   Wherein, X _PBe the power spectrum of radar return:

   $X_p[m,k]={\left|\mathrm{FFT}\right\{x[m,i]w\left[i\right]\left\}\right|}^2={\left|\underset{i=0}{\overset{N-1}{\mathrm{\&Sigma;}}}x[m,i]w\left[i\right]W_N^{\mathrm{nk}}\right|}^2,$ X[m wherein, i] be radar return sampled signal x[n] by fixed step size intercepting, w[i] be window function;

   Step 3, feature extraction: with R (x _v, h _v) discretize obtains R[x _v, h _v], get R[x _v, h _v] in the maximal value of each row, carry out normalization then, intensity distribution image is mapped as an one-dimensional characteristic vector r[x]:

   $r\left[x\right]=\frac{\underset{h}{\max }R[x,h]}{\underset{x,h}{\max }R[x,h]};$

   Step 4, feature screening and compression: adopt Karhunen-Loeve (K-L) to launch, the one-dimensional characteristic vector r[x that step 3 is obtained] project in the new feature space by the ordering of variance size, therefrom choose and obtain reflecting feature and keep the top n proper vector of better classifying quality;

   Karhunen-Loeve (K-L) expansion process is as follows:

   At first computational data intensity distributions proper vector set the covariance matrix ∑ of r}:

   ∑=cov (r)=E[(r-μ) (r-μ) ^T], wherein μ is the population mean vector of feature samples collection,

   Adopt the latent vector of the method compute matrix ∑ of svd (SVD) then:

   $\mathrm{\&Sigma;}={\mathrm{U\&Lambda;}}^{\frac{1}{2}}{V}^T,$ Wherein U, V are orthogonal matrix, are made up of the latent vector of ∑ matrix, and Λ is a diagonal matrix, and element is the eigenvalue of the ∑ matrix arranged from big to small on the diagonal line;

   Utilize latent vector U as orthogonal basis, sample set projected in the new feature space that sorts by the variance size:

   X＝U ^Tr；

   Step 5, by criterion characteristic quantity is compared, finish classification: utilize to obtain normalized optimal weight vector under the Fisher criterion, utilize optimal weight vector that sample is projected to the one-dimensional space and compares the affiliated type of judgement sample with average optimal threshold power;

   The method of two class line styles classification is as follows:

   For the sample vector X of N dimensional feature space, find weight vector w and threshold value power w ₀, make discriminant function:

   G (X)=w ^TX+w ₀Satisfy decision rule:

   

   L wherein ₁, L ₂Be respectively the sample set of two classes;

   The process of determining weight vector w is as follows:

   Definition m _iBe sample average vector, S _iBe sample within class scatter matrix, S _LBe total within class scatter matrix:

   $m_i=\frac{1}{N}\underset{X\&Element;L_i}{\mathrm{\&Sigma;}}X,i=\mathrm{1,2}$

   $S_i=\underset{X\&Element;L_i}{\mathrm{\&Sigma;}}(X-m_i){(X-m_i)}^T,i=\mathrm{1,2}$

   S _L＝S ₁+S ₂

   Then normalized optimal weight vector w is under the Fisher criterion:

   $w=\frac{S_L^{-1}(m_1-m_2)}{\left|\right|S_L^{-1}(m_1-m_2)\left|\right|}$

   Obtain sample to be projected to one-dimensional space Y after the optimal weight vector w, and utilize average method to calculate optimal threshold power Y ₀

   Y _n＝w ^TX _n，n＝1，2，…，N _i

   $m_i=\frac{1}{N_i}\underset{X\&Element;L_i}{\mathrm{\&Sigma;}}Y,i=\mathrm{1,2}$

   $Y_0=\frac{{\tilde{m}}_1+{\tilde{m}}_2}{2}$

   So far, for arbitrary sample X, can use optimal weight vector w to calculate its projection Y, then to Y and Y ₀Size compare, can adjudicate the type of X, obtain vehicle classification.
2. 2, the vehicle type classification method based on single-frequency continuous wave radar according to claim 1, it is characterized in that, described step 5, by criterion characteristic quantity is compared, finish classification: for sample polymorphic type situation, use a plurality of two classification of type devices to carry out pairwise classification respectively, the result according to sorter votes then, adopts sample subpoint in different sorters to arrive the distance D of boundary threshold point _i=| Y _i-Y _0i|, select maximum D _i,, obtain the vehicle classification result as final classification.

Images