Joint Workshop of SCPA 2011 and SaCoNAS 2011

A Solution to the Ill-Conditioned GPS Accuracy Classication Problem: Context Based Classier
Nabil Drawil, Haitham Amar, Otman Basir
Department of Electrical and Computer Engineering University of Waterloo Waterloo, Ontario, Canada N2L 3G1

AbstractGPS localization has been attracting signicant attention recently in many areas. Intelligent transportation systems, navigation systems, road tolling, and collision avoidance systems, are examples of applications that utilize the GPS technology for localization. However, localization accuracy remains a key issue that prevents such applications from delivering on their ultimate promise. The localization accuracy of any GPS system depends heavily on the methodology it uses to compute locations as well as the measurement conditions in its surrounding. The impact of the measurement conditions on the localization accuracy in itself is an intricate ill-conditioned problem due to the incongruent nature of the measurement process. This paper proposes a novel scheme to address localization accuracy. The scheme involves three steps, namely, classify instantaneous GPS accuracy based on measurement conditions, consolidate the consecutive classications of the GPS Accuracy, and enhance the robustness of classifying the GPS accuracy for the following measurements. Real life comparative experiments are conducted to demonstrate the efcacy of the proposed scheme in classifying the GPS accuracy under various measurement conditions. Index TermsGlobal Positioning System, GPS Accuracy, Context Awareness, Measurement Conditions, Neural Network.

I. I NTRODUCTION Global Positioning System (GPS) technology has been used as the de facto standard solution for localization in many Intelligent Transportation Systems (ITS). These systems (i.e., ITS) are aimed to support a wide range of applications [1], such as navigation systems, location based services, accident avoidance and management, location sensitive billing systems. To deliver on their requirements, these applications need a specic localization accuracy. However, GPS based localization lacks the consistency in providing the required accuracy a deciency that many applications can not tolerate, especially mission critical applications. Many pieces of research have been conducted to mitigate the localization error using GPS receiver as a standalone localization system. The majority of the approaches followed in the literature were aimed to integrate one or more localization technique(s) with a GPS receiver in order to constitute a system that works in synergy. For example, in [2], [3], Inertial Navigation System (INS) or other motion measurement system is used along with a GPS or Differential GPS (DGPS). In such systems, the GPS is used to estimate the location unless the satellite signal is lost, then the location is estimated using the motion measurements. It is common, also, to fuse the motion

measurements with the GPS measurements using Kalman Filter (KF) based on the uncertainty in the measurements of the two systems. In [4], [5], vision features extracted from laser scanner or digital camera are fused with digital maps and GPS/INS localization system. Moreover, in [6], GPS/INS localization system is fused with radio location techniques to enhance the cooperation among different vehicles trying to correct their uncertain location estimate. Although these localization techniques have demonstrated signicant performance improvement, there remain situations that give rise to degraded localization accuracy. In other words, localization techniques can not guarantee a consistent level of accuracy as a performance measure. The reason for this degradation is that the amount of uncertainty in the location information sources changes in a highly stochastic manner. The focus of this paper is on studying the change in the accuracy performance of a localization system, namely a stand-alone GPS receiver. The selection of such a localization device is based on the fact that GPS is a basic building block in most of the localization techniques in the ITS. Indeed, any degradation in the GPS localization accuracy performance is reected on the performance of a wide range of localization systems. The engineering knowledge of a localization device is crucial in evaluating its accuracy. Therefore, different parameters are extracted from a GPS receiver; the effect of which on the localization error is studied. Normally, the change in these parameters is inuenced by the measurement conditions. The recognition of the measurement conditions that have an impact on the performance of a GPS accuracy is a key issue to evaluate the quality of its location estimate. According to the knowledge of the authors, only a few parameters have been studied in the literature in terms of their relationship to the change in the GPS accuracy performance. These parameters are signal to noise ratio (SNR) of the received signal, dilution of precision (DOP), and number of observed Satellites (cf., [7], [8], [9]). Nevertheless, to the best of our knowledge, GPS accuracy has not been addressed as a pattern recognition or classication problem. In this paper, the parameters extracted from a GPS receiver are used to construct a feature space in which measurements appear as feature points. The GPS accuracy is classied based on the trend formed by a set of these feature points. In contrast to the research work conducted in this area, the approach

978-1-4673-0040-7/11/$26.00 2011 IEEE

1077

taken in this paper starts by considering more parameters and providing a description of the GPS accuracy for each location estimate. The application utilizing the GPS estimate can then make a decision whether to use, ignore, or even improve the GPS estimate by fusing it with other location estimates obtained by other modalities. In this paper, furthermore, a novel classication algorithm is proposed to enhance the classication of the performance of a GPS receiver. The proposed algorithm maintains the fact that instantaneous measurements provide ill-conditioned evidence about the surrounding environment, and complements this evidence with context recent knowledge. The remainder of this paper is organized as follows. Section II describes the components of the GPS error, the construction of the feature space, and the analysis of the GPS accuracies. Section III presents the proposed classication algorithm. Section IV explains the real life experimental work and also gives the results of the proposed classication algorithm. Finally Section V summarizes the paper and provides concluding remarks. II. E RRORS IN GPS L OCATION E STIMATE As the signals of the GPS satellites travel to the GPS receiver, several error components accumulate in the pseudorange code code , and thus constitute the GPS localization error. These error components are summarized in the following equation code = r + eph + iono + trop cT + mp + vrcvr (1)

(NMEA) sentences for every location estimate. Typically, these parameters are function in the observation conditions. The extracted parameters include number of satellites used in the localization process, the DOP, the mean and standard deviation of the SNR measures of the satellites used in the localization, and the speed of the vehicle reported by the GPS. In the remaining part of the paper, the values of the parameters will represent the features and a group of features pertaining to a location estimate represents a pattern, P Rd , where d is the number of features. For each measurement, the error in the location estimate is computed. The error is dened as the lateral (cross-road direction) distance between the GPS location estimate and the ground truth location, Figure 1.
# # # # # # # #

# # # ## ## # # # # # # # # # # # # ## ## # # # # # # ### # #

Vehicle Location Estimate Localization Error Road Traced Roads Digital Map
0.06 0 0.06 0.12 Kilometers

where r signies the range between the receiver and the satellite, and eph signies the satellite ephemeris error. The ephemeris error is caused by the difference between the actual GPS satellite position and the anticipated position. iono and trop signify the ionospherical and tropospherical errors, respectively, and they are caused by reduction in the speed of the satellite signal. The signal originally travels at the speed of light, c, but the speed gets reduced when crossing the ionosphere and troposphere layers. The receiver clock error is signied by T . The mp signies the multipath error. The multipath error is caused by the delay of the signal arrival due to the reection off the building surfaces in urban canyons. The error component vrcvr signies the receiver measurement noise which is device dependent. Rankin [10] has reported that the multipath and device-noise error components dominate the error in code . Moreover, the multipath error has unknown distributions and bias in some situations, as stated by Syed and Cannon in [11]. In fact, the GPS localization error ranges from few meters in open sky environments to over 80 meters in urban canyons. The focus of this paper is to classify the observation conditions and to infer the localization accuracy accordingly. To accomplish this task, a real life GPS data is collected for a vehicle travels in different environment conditions (e.g., highway, suburban, urban canyons) so as to obtain a comprehensive set of data. A number of parameters are obtained from the GPS receiver, using the National Marine Electronics Association

Fig. 1.

GPS Localization Error.

This error computation ignores error components along the road direction. However, as seen in Figure 1, this component is small compared to the amount of error nominally experienced orthogonal to the vehicle direction. The error computed for each sample is mapped to an accuracy class (i.e, high, mild, low accuracy). In what follows we describe the approach we used to analyze the collected measurement samples. A. Localization Error Analysis Principal Component Analysis (PCA) is applied on the collected patterns Pk to analyze the behavior of the features with respect to the localization error. As a rule of thumb PCA is able to maximize the variation of the features, as such it is possible to detect the most indicative features. It can be seen in Figure 2, which shows the PCA in three dimensions (i.e., the rst three principal components), that the data clearly has three clusters. PCA transformation shows that the speed, signal to noise ratio mean (SN R ), and signal to noise ratio standard deviation (SN R ) are the main features that constitute the three clusters. A K-means algorithm is used to label the samples belong to the three clusters in the PCA feature space. Now, if one can draw a relationship between these clusters and the localization error, it would be possible to infer the localization accuracy using the aforementioned PCA feature space. Without loss

1078

TABLE II F EATURES AND LOCALIZATION ERROR RELATIONSHIPS .

Accuracy Band Class-1 Feature # of Sat DOP SN R and SN R Speed # of Sat DOP SN R Speed # of Sat DOP SN R Speed Range # of Sat > 10 DOP 1.0 >35 and SN R <12. Mid to high # of Sat 4 DOP 6 SN R < 35 Low to Mid # of Sat 4 DOP > 6 SN R < 35 Low to Mid

SN R

Class-2

Class-3

Fig. 2.

Data projection on the rst three Principal Components.

of generality, the range of the localization error is divided into three bands. As seen in Table I, these bands constitute our accuracy classes: Class-1 signies high accuracy, Class2 signies mild accuracy, and Class-3 signies low accuracy. These bands are application dependent and can take various spans over the error range.
TABLE I T HREE CLASSES OF LOCALIZATION ACCURACY. Class-1 Error range 0-10 (m) Class-2 10-20 (m) Class-3 greater than 20 (m)

Hence, FFNNET is chosen as a conventional classier for this research work. Nonetheless, the GPS accuracy changes within a volume of time and space. Therefore, a use of a single pattern classier may pose the threat of arriving at ill-conditioned classication, in which one measurement belongs to more than one class. Indeed, a comprehension of historical changes in the environment can help realizing the current change of the GPS accuracy within the environment. In the next section, we propose a solution to the ill-conditioned GPS accuracy classication process. III. C ONTEXT BASED C LASSIFIER (CBC) As stated in the previous section, it is difcult to utilize a conventional classier to distinguish between our three classes at once. To realize our goal of distinguishing the three classes, we introduce Context Based Classier (CBC) algorithm. This algorithm is used to enhance the performance of a classier by providing a recent historical knowledge about the surrounding environment as it is shown below. Consider pk patterns where k = 1, . . . , K , pk Rd , d is the number of features in a pattern, and K is the number of these patterns. These patterns belong to M classes labeled numerically by the set C = {1, . . . , M }. The idea is to use a conventional classier, h1 (pk ), in order to generate preliminary classications for the patterns pk . A successive k ) are used to create f(k,L) R, output of h1 (i.e., y yk , . . . , y kL ) f(k,L) = f unction( (2) where y i C , i = k, . . . , k L. This feature, f(k,L) , is fed along with pk to another conpk ), where p k = (pk , f(k,L) ) and ventional classier, h2 ( p k Rd+1 . Basically, the feature f(k,L) is introduced to h2 to indicate the likelihood of belonging to certain class (i.e., classcandidate). Therefore, f(k,L) diminishes the similarity between classes in the new space Rd+1 . The computation of f(k,L) is illustrated below. A class-candidate is obtained using a historical window buffer, W n(k,L) , of length L containing the previous clasyk , . . . , y kL )). The sications of h1 (i.e., W n(k,L) = ( class-candidate value signied by f(k,L) shall indicate two pieces of information: 1) the class to which the surrounding

Having the above mentioned sets of classes and clusters, we attempt to map a relationship between the accuracy classes and the clusters by associating the location errors within the bands to the clusters in the PCA feature space. As can be seen in Figure 2, most of Class-1 can be separated from the other two classes. Nevertheless, Class-2 and Class3 are strongly overlapped and mixed with some patterns from Class-1. This observation leads to the fact that the PCA transformation is not sufcient to fully separate the three classes due to the high degree of similarity between Class-2 and Class-3 in the feature space. Therefore, the three accuracy bands have been investigated in terms of their feature trends respectively as depicted in Table II. It is concluded that Class-2 and Class-3 are similar in terms of the features deemed to be the most indicative (i.e., speed, SN R , SN R ), as stated above. Classifying such distributed feature points is a challenge that is impossible to be resolved using traditional classiers, such as Linear Discriminant Analysis methods. Kernel based classication methods such as Support Vector Classiers are expected to provided decent performance in such a space conguration; nevertheless, the computation of the kernel based classication is very expensive a deciency that can not be tolerated by many critical mission applications. Feed Forward Neural Network (FFNNet) is known for its practicability in embedded systems, and able to recognize linearly inseparable classes, when sufcient training set is available.

1079

environment is pertaining, and 2) the uncertainty about this classication. The class-candidate is chosen according to the majority of W n(k,L) entries, and the uncertainty is computed using entropy (i.e., H ), as in Equation 3.
M

HW n(k,L) =
i=1

P ri (W n(k,L) ) logM P ri (W n(k,L) ) (3)

where P ri (W n(k,L) ) is the probability that the current, k , reading belongs to class ci , i {1, . . . , M }. P ri (W n(k,L) )s are computed statistically using W n(k,L) . Equation 4 is used to compute f(k,L) given the distribution of the class-candidate and entropy value. A = arg max P ri (W n(k,L) ), i {1, . . . , M }
i

f(k,L) =

B = arg max P ri (W n(k,L) ), i {1, . . . , M } A i HW n(k,L) when A B 2 A A + (4)

HW n(k,L) 2

when A < B

Based on equations 3 and 4, the range of f(k,L) is [1, M ]. The lower the uncertainty of the class-candidate, the closer the feature, f(k,L) , value gets to the class label and vice versa.
CBC Output:

Algorithm 1 Pseudocode for CBC Input: Testing patterns pk s; Output: Classication of the input y k s; k is the ultimate output /*Note: y k is the output of h1 and y of CBC*/ /*The rst L patterns are classied using only h1 */ for k = 1 to L do y k = h1 (pk ); ShiftRight(Buffer(y )); k ; y k = y end for /*Start feeding h2 */ loop k + +; y k = h1 (pk ); ShiftRight(Buffer(y )); /*Calculate P ri s, HW n(k,L) , and f(k,L) as in equations 3 and 4*/ /*Construct pattern for h2 */ p k = (pk , f(k,L) ); pk ) ; y k = h2 ( end loop

k y
 yk ,..., yk  L 

classiers used in the classication techniques are provided. Comparative results and discussion conclude this section. A. Experiment Setup A roving vehicle equipped with a standard GPS receiver (SiRFstarIII Evaluation Receiver) has been used to collect the measurements for our experiment. A Lenovo SL400 Laptop connected to the GPS receiver over a USB port has been used to store the NMEA sentences to be processed later ofine. The experiment has been carried out over an area with a high degree of variety of measurement conditions. Open sky environment (i.e., high accuracy conditions) has been tested over the highway 401 between Waterloo and Toronto in Ontario. The low and mild accuracy regions have been tested in the suburbs and downtown of Toronto-Ontario, Figure 4. B. Classiers Implementation

Buffer Window of Size L (Wn(k,L))

Wn k , L 

yk
Conventional Classifier (h1) Conventional Classifier (h2) ClassCandidate f ( k , L ) Computation

CBC Input: pk

 f1k ,..., fdk 

Fig. 3.

Context Based Classier (CBC) schema.

Accordingly, the new set of patterns (i.e., p k ) is constructed k k , . . . , fd , f(k,L) , and used to train the second as p k = f1 conventional classier, h2 . Figure 3 shows the whole picture of CBC, in which two conventional classiers collaborate to utilize f(k,L) . Detailed procedures performed in testing phase of CBC are depicted in Algorithm 1. In the next section, experimental result are given based on real-life data by both a conventional classier and CBC. IV. E XPERIMENTAL WORK This section presents the implementation of the algorithm proposed in this paper. First, the setup of the real-life experiment conducted is explained. Then the implementation of the

The cross-validation criterion is the testing approach for validating the proposed classication technique in this paper. In order to balance the data set among the three classes, equal number of patterns has been collected for each class. The data set is divided into training data set and testing data set with percentages of 75% and 25%, respectively, for each class. Without loss of generality, the classiers, h1 and h2 , used in this work have be chosen to be FFNNet with backpropagation learning algorithm due to its ability to classify linearly inseparable patterns. Nevertheless, the strong overlap among the classes causes the FFNNet to misclassify some patterns as it is impossible for the FFNNet to avoid such misclassications. The goal of the experiment is to show how CBC strongly diminishes these misclassications.

1080

Fig. 4.

Real life experiment eld.

Each FFNNet consists of three layers (i.e., input, hidden, and output layers). The number of nodes in the input layer is equal to the number of features. The size of the hidden layer (network complexity) has been determined based on optimizing the testing error rate using cross-validation. As seen in Figure 5 It is evident that the optimal number of the nodes in the hidden layer is between 22 and 25. Increasing the complexity will cause overtting which degrades the performance of the FFNNet. The number of nodes in the output layer is equal to the number of classes. The activation function in the input layer is linear, while it is tansig in the hidden and output layer.
24.8 24.6 24.4

percentage of patterns belong to certain class and classied as belonging to other classes. The forth column shows the total error rate that signies the percentage of patterns misclassied in the testing data set. It can be seen that Class-2 has the largest error rate due to the strong similarity of this class with the others. CBC shows a signicant improvement, about 14.5%, in the error rate of Class-2 and , also, decreases the error rate of the other classes. The improvement in the total error rate is 6.67%.
TABLE III E RROR RATE FOR THE THREE LEVEL ACCURACY CLASSIFICATION . Class-1 FFNNet CBC 24.50% 20.50% Class-2 52.50% 38.00% Class-3 13.00% 11.50% Total 30.00% 23.33%

Performance

24.2 24 23.8 23.6 23.4 23.2 23 10

The confusion matrix of the two classiers are presented in tables IV and V. These matrices details the classication results, where each row represents certain class and each column represents the percentages of patterns classied as belonging to certain class. The increase in the values of the diagonal entries in Table V shows the improvement achieved using CBC.
15 20 25 30 35 40 45

Neural Network Complexity

TABLE IV C ONFUSION M ATRIX FOR FFNN ET CLASSIFIER Classication Class-2 Class-3 14.50% 47.50% 13.00% 10.00% 36.00% 87.00%

Fig. 5. Optimizing the Neural Network Complexity. The optimal complexity produces the minimum testing error rate. Class-1 Actual Class-2 Class-3

Class-1 75.50% 16.60% 00.00%

C. Comparative results The comparison here is between the FFNNet and the CBC performacne in classifying the patterns of the testing data set into three classes (i.e., high, mild, and low accuracy). The testing data set of size 200 patterns per class is used to perform a comparison in terms of classication error rates. The whole data set, patterns collected in trip from Waterloo to downtown of Toronto, are tested to produce visual comparisons in different environment conditions. 1) Error rate comparison: Table III shows in the rst three columns the error rate per each class which signies the

2) Visual Comparison: Since the performance of the FFNNet and the CBC are comparable over the highway 401 and accurate, the focus here is on certain scenarios where misclassication is common. Table VI shows two snapshots taken from the GPS location estimates. The st row shows the labeled data, where the second and third rows show the performance of the FFNNet and CBC respectively. The rst snapshot, in the rst column, is taken while the vehicle is going through an underpass. In such conditions

1081

TABLE VI ACCURACY CLASSIFICATION IN DIFFERENT MEASUREMENT CONDITIONS Underpass in highway regions High rise building area
$ $ $ $ $ $ $$$ $ $ $ $ $ $ $$ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $$ $ $ $ $ $ $ $ $ $ $$ $ $ $ $ $ $ $ Labeled Measurements $$ $ $ $ $ $ $ $ $ $$ $ $ $ $ $ $ $ $$ $ $ $ $ $ $ $$ $ $ # Class-1 $ $ $$ $ $ $ % % % $ $ $ % $ % % % # % $$ % Class-2 $ % $ $$ $ $ $ $ # $ $ %%% $ % $ $ $ $ % %% $$$ % $$ $ $ $ % % $ $ % # $ $ % $ # $ $ # $ $ # % $ $ # $ % $ % %% $ $ # % # # $ $ $ $ $ $ # ## # $% % % %% # # # $ Class-3 # # # % %%% # $ # # ## # % % %% $ # $ $ % $ % % % $ $ % $ % $ $ $ $ # Ground Truth Path (Downtown Toronto) $ # $ %$ $ % % $ % $ % % % % % % % % % % # % % % % % % % Roads Digital Maps $ % $ $ % $ $ % $$ $ $ $ $ $ $ $ $ % $ $ $ $$$$ $ # % $ # %%% $ %% % # % % ## $ % % % % # ### # % # #% # # %$ % % % % % % $ # $ # $ $ 40 0 40 80 Meters $ $ $ # $ $ $ $ $

Labeled patterns

Labeled Measurements # Class-1 % Class-2 $ Class-3 # Ground Truth Path Roads Digital Maps
50 0 50 100 Meters

## # # # ### # # # # # # # #

##

FFNNet classication

FFNNet # Class-1 % Class-2 $ Class-3 # Ground Truth Path Roads Digital Maps
50 0 50 100 Meters

%$ $ $% % % % %% % $ # # # % # #

FFNNet # Class-1 % Class-2 $ Class-3 Ground Truth Path (Downtown Toronto) Roads Digital Maps $ $$ $
40 0 40 80 Meters

$ $$ $ $ $ $$$ $ $ $ $ $ $ $$ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $$ $ $ % $ % $ $ $ $ % $$ $ $ % % $ $ $$ $ $ $ $ $ $ $ $ $$ $ $ $ $ $ $ $ $$ $ $ $ $ $ $ $ $ % $ $ $ $$ $ $ $ $ $ $ % $$ $ $ $ $ $ $ $ $ $ $ $$ $ $ $ $ $ $ $$$$ $ $ $ $ % % $ $$ $$% $$ $ % $ % $ $ $ $ $ $ $ $ % $% % $ $ % $ $ $ $$ $ $ $ % % $ $ $ $ $ $ $ $ $ $$ $ %%%$ % % $ $ $$ $$ $ $ $ $$ $ % $% $ $ % $ % $ % $% $ % $ $ $ $ $ $ $ % $ $ % % $ $ % $ % $ $ % $ $ $ $ $ $$ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ % $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $$$$ $ $ $ $$ $$$ $$ $ $ $ $ $ $$ $ $ $ $ $ $$$ $ $ $ $$ $ $$ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $

CBC CBC

CBC classication
50

# % $

Class-1 Class-2 Class-3 # Ground Truth Path Roads Digital Maps

0 50 100 Meters

## # # # ### # # # # # # # #

#$

# % $

40

% $% % $ $ $ $$% $ $ $ $ $ $ $$ $ $ $ $ $ $ $ $ % $ $ % $ $ $ $ $ $ % % $ $ $ $$ $ % $ $ $ $ $ $ $ $ %% $ $ $ $ $ $ $ $% $ $ $ $ % $ % $% $ $ $ % $ $ $ $$ $ $ $ $ % $ $$ $ $ Class-1 $ $ % %$ $ % $ $ % % % $ $ % $% $ % $ % $ % $ $ Class-2 $ $% % $ $ %% % $ $ $%% % $ $ $ $ $ % $$$ % %% $$ $ % $ $ % % $ $ $ % % $ $ $ $ $ % % $ $ % $ % $ $ $ % $ $ $ $ $ %% $ $ $ $ $ % % % $$ $$$ % % $ $ $$$ $ % $ %% $ % %% % Class-3 % $ $ % $ $ $ $$ % $ $ $ %% % $ % $ $ % $ % % % $ $ % $ % $ $ $ $ $ $ Ground Truth Path (Downtown Toronto) $ % $ $ $ $ $ $ $ $ % $ $ $ % % % % % % % $ $ % % % % % % $ Roads Digital Maps % $ % $ % $ % % $ $ $ $ $ $ $ % $ $ $ $ $ $ $$%$ % $ $ % %%% $ %% $ % $ % $% $ % % % % % % %% % % $ $ $$ $ $ $$ % % $ $ % $ % $ $ % $ % $ $ 0 40 80 Meters % $ $ % $ $ $ $ $

TABLE V C ONFUSION M ATRIX FOR CBC CLASSIFIER Classication Class-2 Class-3 13.00% 62.00% 11.50% 07.50% 22.50% 88.50%

Class-1 Class-1 Actual Class-2 Class-3 79.50% 15.50% 00.00%

conducted to show the signicant improvement achieved when using CBC. Using CBC it is possible in most of the cases to determine the accuracy of a GPS receiver even in chaos measurement conditions. Yet, the condence level of the decision made about the accuracy remains questioned and open for research. R EFERENCES
[1] A. Boukerche, H. Oliveira, E. Nakamura, and A. Loureiro, Vehicular Ad hoc Networks: A New Challenge for Localization-Based Systems, Computer Communications, vol. 31, no. 12, pp. 28382849, 2008. [2] W. Kao, Integration of GPS and Dead-Reckoning Navigation Systems, Vehicle Navigation and Information Systems Conference, 1991, vol. 2, pp. 635643, Oct. 1991. [3] Q. Honghui and J. B. Moore, Direct Kalman Filtering Approach for GPS/INS Integration, IEEE Transactions on Aerospace and Electronic Systems, vol. 38, no. 2, pp. 687 693, 2002. [4] A. Rae and O. Basir, A Framework for Visual Position Estimation for Motor Vehicles, 4th Workshop on Positioning, Navigation and Communication, pp. 223228, Mar. 2007. [5] M. Jabbour, P. Bonnifait, and V. Cherfaoui, Enhanced Local Maps in a GIS for a Precise Localisation in Urban Areas, IEEE Intelligent Transportation Systems Conference, pp. 468473, Sep. 2006. [6] N. Drawil and O. Basir, Intervehicle-Communication-Assisted Localization, Intelligent Transportation Systems, IEEE Transactions on, vol. 11, no. 3, pp. 678691, 2010. [7] M. Mosavi, K. Mohammadi, and M. Refan, Fuzzy Processing on GPS Data to Improve Positioning Accuracy, before and after S/A is Turned Off, in The Asian GPS Conference, 2002, pp. 117120. [8] T.-H. Chang, L.-S. Wang, and F.-R. Chang, A solution to the illconditioned gps positioning problem in an urban environment, Intelligent Transportation Systems, IEEE Transactions on, vol. 10, no. 1, pp. 135 145, Mar. 2009. [9] J. Meguro, T. Murata, J. Takiguchi, Y. Amano, and T. Hashizume, GPS Multipath Mitigation for Urban Area using Omnidirectional Infrared Camera, Intelligent Transportation Systems, IEEE Transactions on, vol. 10, no. 1, pp. 2230, 2009. [10] J. Rankin, An error model for sensor simulation gps and differential gps, Position Location and Navigation Symposium, vol. 11, no. 15, pp. 260266, Apr 1994. [11] S. Syed and M. Cannon, Fuzzy Logic-based Map Matching Algorithm for Vehicle Navigation System in Urban Canyons, in ION National Technical Meeting, San Diego, CA, vol. 1, 2004, pp. 2628.

the location estimate is accurate, but the features are similar to those of high rise building areas where the localization accuracy is low or mild. Therefore, The FFNNet is misled by the features and fails to classify the accuracy. On the other hand, CBC succeeds in classifying the accuracy. The second snapshot, in the second column, is taken while the vehicle is traveling through a high rise buildings area. In such conditions the accuracy of the location estimate is not consistent, but it is often low. In such environments the GPS may produce location estimates with small location error, but the features are totally different from those of high accuracy. As seen in Table VI the FFNNet classies most of the patterns as low accurate; however, some patterns have mild accuracy and others are accurate. On the other hand, CBC succeeds to capture most of the mild accurate patterns and considers the accurate patterns as mild accurate. V. C ONCLUSION This paper deals with accuracy classication based on measurement conditions. The localization error in the GPS estimate is analyzed and different measurement conditions are studied. A novel classication technique, context based classier (CBC), is proposed as a solution for the ill-conditioned GPS accuracy classication problem. A real life experiment was

1082