Classification of Vehicles Using Magnetic Dipole Model

Transcription

1 Classification of Vehicles Using Magnetic Dipole Model PrateekGV,Rajkumar V, Nijil Kand K.V.S.Hari and ssplab Statistical Signal Processing Laboratory, Department of ECE Indian Institute of Science, Bangalore, India. IEEE TENCON 2012, Cebu, Philippines 21 st November2012 Funded by DIT-ASTEC Wireless Sensor Project, Department of Information Technology, Ministry of Communications & Information Technology, Govt. of India.

2 MOTIVATION FOR CLASSIFICATION OF VEHICLES One important requirement for a traffic management systemis thecapability todetectthepresenceofavehicle and typeofavehicle (car, bus,truck,etc). Basedonsuch detection, statistics such as vehicle count trafficflow speed occupancy Induction loop and Video-Image are used most widely technologies but they have a lot of disadvantages. Inductionloops arebig in sizewith difficultyin maintenance. Video-Imagebasedsensor arecostly with big influence of external light conditions.

3 CLASSIFICATION OF VEHICLES USING MAGNETIC SIGNATURES Passivemagnetometers that arecapable ofsensingthe magnetic field can be used. Themotes having these sensorsmounted onthem can be programmedwith avehicle detectionalgorithm High level of flexibility in their deployment configuration and costs less. a. magnetic perturbations c. sensor readings b. AMR sensor Anisotropic Magnetoresistive(AMR) sensors detect the distortions of theearth smagnetic field, whichis assumed to beuniformover awide areaonthe scale ofkilometers. S.Y. Cheungand P. Varaiya,Trafficsurveillance bywireless sensornetworks,researchnote, University of California, Berkeley,Jan UCB-ITS-PRR pdf.

4 DATA COLLECTION Data is collected using two different mechanism. (a) Remote Controlled Car (b) Skate Board Paths acrosswhich the HMC1502 sensormounted on atelosb wirelessmote placed in a fibercasing, witheither a remote control car setup or skate board setup, was moved

5 DATABASE - VEHICLE MAGNETIC SIGNATURES VehicleMagneticSignatureDatabase groupedbasedonthe lengthofthe car Car-type Type 1 Type 2 Type 3 Type 4 Car Length (in meters) ( ) ( ) ( ) (>4.5) Type of 1 800(8) 11 Corsa(2) 3 Accent(1) 6 Civic(1) Car(n), 1 Alto(2) 3 i20(1) 2 Cielo(1) 8 Corolla(1) where n 2 Matiz(3) 5 Figo(2) 6 City(4) 3 Elentra(2) represents 3 Santro(5) 3 GetZ(2) 12 Vento(1) 8 Innova(2) number of 1 Omni(6) 3 i10(4) 1 SX4(2) 7 Linea(1) datasets 9 Spark(1) 4 Indica(6) 3 Verna(1) 3 Sonata(1) 4 Nano(2) 7 Palio(1) 1 Esteem(2) 10 Octiva(1) 1 WagonR(4) 1 Swift(2) 4 Indigo(2) 10 Laura(1) Cars=42 1 Estillo(3) 1 Zen(2) 1 Dzire(1) Sets=89 9 Beat(2) 3 Ritz(1) 4 Sumo(1) 13 Reva(1) 5 Fiesta(1) 6 Petra(1) 14 Logan(1) Number of Datasets Indicates thecarmanufacturer 1 -Maruti Suzuki; 2 -Daewoo; 3 -Hyundai; 4 -Tata Motors; 5 -Ford; 6 -Honda; 7 -Fiat; 8 -Toyota; 9 -Chevrolet; 10 -Skoda; 11 -Opel; 12 -Volkswagon; 13 -Mahindra; 14 -Renault. A.S. Bhat, A. K. Deshpande,K. G. Deshpande, and K. V.S. Hari, Vehicle detection and classification using magnetometer- data acquisition, tech. rep., 2011.

6 SAMPLE MAGNETIC SIGNATURES 1800 Type 1(Alto) y axis reading Type 1 Alto Type 2(Indica) y axis reading Type 2 Indica M y signal amplitude M y signal amplitude Sample Index (e) Y-axisreading fortype1-maruti Alto Sample Index (f) Y-axis reading fortype2-tata Indica 1900 Type 3(SX 4) y axis reading Type 3 SX Type 4(Sonata) y axis reading Type 4 Sonata M y signal amplitude M y signal amplitude Sample Index (g) Y-axis readingfortype3 -Maruti SX Sample Index (h) Y-axis reading fortype4-hyundai Sonata The Y-axis trajectories obtained using HMC1502 magnetometer of cars belonging to different types( Length of Car(inm)-( ) Type1; ( ) Type2; ( ) Type3; (>4.5) Type4)areshown.

7 Problem Statement: To classify vehicles using magnetic signatures obtained from passive magnetometers. Steps involved in solving Data Modeling of magnetic signature Extraction of feature vector from the magnetic signature. Use classification techniques and study the performance of the classifier.

8 DATA MODEL - MAGNETIC DIPOLE MODEL A vehiclecan be modeledas anarrayof dipoles. Illustration of a Magnetic Dipole Model for a Vehicle. m(i)where,i {1,...,M}representsmagnetic dipolemoments, X(j) where,j {1,...,M 1}isthe separationbetween adjacent dipoles, Y and Zarethe offsets,v 0 bethe velocityofthe vehicleand r 0 be distance of m(1) fromthe sensorplaced atthe origin. N. Wahlstrom, J.Callmer, and F. Gustafsson, Magnetometersfortracking metallic targets, in Information Fusion(FUSION), 2010

9 DATA MODEL - MAGNETIC DIPOLE MODEL If the distance fromthe objectis largeincomparisonwithits characteristic length, the inducedmagnetic field B(r,m) at position r = [x,y,z] T relativetothe objectcan be describedas amagnetic dipole fieldisgivenas µ 0 3(r m)r r 2 m B(r,m) = (1) 4π r 5 where B(r,m) = [B (x) (r,m),b (y) (r,m),b (z) (r,m)] T, m = [m (x),m (y),m (z) ] T isthe magnetic dipolemoment, r = r 2 isthe L 2 -Normand (r m) isthe scalardotproductof the two vectors. Substituting r = [x,y,z] T and m = [m (x),m (y),m (z) ] T inequation(1) gives the following B (x) (r,m)= µ0 (3x 2 r 2 )m (x) +3xym (y) +3xzm (z) (2) 4π r 5 N. Wahlstrom, J.Callmer, and F. Gustafsson, Magnetometersfortracking metallic targets, in Information Fusion(FUSION), 2010

10 SENSOR INDEPENDENT APPROACH Inthe signalprocessingframework,asensorcan be modeledas a time-invariant system y k = f(r k,m k )+e k (3) = µ0 3(r k m k )r k r 2 km k +e 4π r 5 k (4) k The number ofparametersto be estimatedforanm-dipolemodelis 4M+1 p = [m(i) T, X(j), Y, Z] T The vehicleis assumedto move paralleltothe X-axis,the only time varyingcomponent inr k isx k f(r k,m k ) = f(x k,p) (5) Let ˆpbe the estimate ofp. Then,the Non-linearLeastSquares(NLS) cost function gives the following where, V(p) = ˆp = arg min p V(p) (6) N [y k f(x k,p)] T [y k f(x k,p)] (7) k=1

11 MAGNETIC DIPOLE MOMENTS AND DIPOLE SEPARATION ALGORITHM (MDMS ALGORITHM) Input: Smoothed VehicleMagnetic Signature-a N 1 Input: The number of magnetic dipoles- M 1: Subtract everyk th,k {1,...,N}samplewith the meanof firstn/10 samplesofa N 1 2: Getthe Data Modela k = f(r k,m k )+e k 3: V(p) = N [a k f (x k,p)] T [a k f (x k,p)], pbe the parametersto k=1 estimated and p = [m(i) T, X(j), Y, Z] T 4: Estimate p, ˆp = argmin p V(p) 5: NormalizedMagnetic Moments m(i) = m(i) m(i) 2 Output: NormalizedMagnetic DipoleMoments m(i),i {1,...,M}; Separationbetweenadjacentdipoles X(j),j {1,...,M 1}and RMSE where, (RMSE) 2 = 1 N N k=1 [y k f (x k, ˆp)] T [y k f (x k, ˆp,)]

12 SIMULATION RESULTS B (y) Signal Amplitude Three Dipole Model Fitting; M = 3 Data Fit Time (s); F = 100Hz s B (y) Signal Amplitude Four Dipole Model Fitting; M = 4 Data Fit Time (s); F s = 100Hz (i) 3-Dipole Model curve fit fora Tata Indica magnetic reading. (j) 4-Dipole Model curvefit foratata Indica magnetic reading. Sample curve fitting plots for measurements corresponding to a Tata Indica car using MDMS algorithm. M {3,4}. Sampling Frequency, F s = 100Hz. Theerrorin the fit decreasesas numberof dipoles increases. Location: IISc Campus m-dipole Model with Dipole Separation, Dipole Moments and RMSE for a Tata Indica Car s Magnetic Signature M-Dipole X(j) m(i) = m(i) m(i) 2 RMSE 3-Dipole m(1) = ( 0.77, +0.33, 0.52) m(2) = (+0.26, 0.18, +0.94) m(3) = ( 0.71, 0.19, 0.67) Dipole m(1) = (+0.79, +0.29, 0.52) m(2) = ( 0.43, 0.06, +0.89) m(3) = (+0.35, +0.93, 0.05) m(4) = ( 0.34, 0.93, +0.04) 12.2

13 COMPUTATION COMPLEXITY The computational complexity of the NLS cost function using MATLAB function lsqcurvefitis O{(4M+1) 3 }. As the number ofdipoles increasesby1, the number ofparametersto be estimatedincreasesby4 and so does the complexity. Number of Parameters and RMSE for Available Datasets M-DipoleModel Size ofp=(4m+1) 1 RMSE 3-Dipole Dipole Inorderto checkthe variation ofrmse asthe numberofdipoles M increases, wecalculate theaveragermse forall the datasets D across differentvalues of M. RMSE = 1 D RMSE i (8) D i=1

14 EXISTING ALGORITHMS FOR CLASSIFICATION 150 Z axis plot for a Tata Indica Vehicle from Sensor S3 200 Z axis plot for a Tata Indica Vehicle from Sensor S3 Amplitude Amplitude Sample Index. Sampling Frequency F = 100Hz s Average Bar plot for a Tata Indica Vehicle from Sensor S3 Amplitude c Signal Hill Pattern Sample Index. Sampling Frequency F s = 100Hz Sample Index. Sampling Frequency F s = 100Hz Average Bar Sample Index. Sampling Frequency F = 100Hz s (k) Average-Bar Transform: Here the vehicle signature vectoroflength N, is divided into S sub-vectors. The mean value of each sub-vector is calculated and the obtained values for S sub-vector is the featurevector. Thevalueof Sis fixedforall classes of vehicles. (l) Hill-Pattern Transform: This method transforms thesignal into asequenceof{+1, 1}and without losing much information. This extracts the pattern of peaks and valleys (local maxima and minima) of theinput signal. Thesequenceof{+1, 1}is used as a feature vector. S.Y. Cheungand P. Varaiya,Trafficsurveillance bywireless sensornetworks,researchnote, University of California, Berkeley,Jan UCB-ITS-PRR pdf.

15 CLASSIFICATION METRIC Weassume L tr and L ts to be the number of trainingand testingdatasets picked. Wedefine the correctrateof classification, C R as follows C R = 1 I I i=1 Ω i L ts (9) whereω i is the number ofvehiclesclassifiedcorrectlyamong L ts number ofcars inthe i th iterationand the total numberof iterationsisi.

16 CLASSIFICATION USING SVM ThegoalofaSupportVectorMachine(SVM) is toproduce a model(based on the training data) which predicts the targetvalue of thetestdatagivenonlythetestdata attributes. PercentageofC R fortype1vs Type4Car foraveragebar, Hill TransformandMDMS Algorithm Datasets Feature Extraction Algorithms (L tr,l ts ) MDMS Algorithm Average Bar Hill Transform 3-DM 3-DM Algorithm Algorithm m X (70,44) (80,34) (90,24) PercentageofC R fortype2vs Type3Car foraveragebar, Hill TransformandMDMS Algorithm Datasets Feature Extraction Algorithms MDMS Algorithm Average Bar Hill Transform 3-DM 3-DM (L tr,l ts ) Algorithm Algorithm m X (70,50) (80,40) (90,24)

17 Thank You