Departamento de Engenharia de Transportes Universidade Federal do Ceará Department of Civil and Environmental Engineering The University of Tennessee Knoxville, TN, EUA ! Department of Industrial and Systems Engineering Rutgers University, NJ, EUA Freeway automatic incident detection (AID) has been extensively investigated over the last four decades. However, a recent nationwide survey in the United States concluded that the implementation of AID algorithms in traffic management centers is still very limited. The main reasons for this discrepancy are high false alarm rates and calibration complexity. This paper presents a self+learning, transferable algorithm that requires no calibration. The dynamic thresholds of the proposed algorithm are based on historical data of traffic, thus accounting for typical variations of traffic throughout the day to reduce false alarms rate. The proposed model performed better than existing algorithms found in the literature. " Detecção automática de incidentes em rodovias tem sido extensivamente investigada nas últimas quatro décadas. Contudo, uma pesquisa recente realizada nos EUA concluiu que a implantação desses algoritmos em centros de controle de tráfego ainda é bastante limitada. A principal razão para esta discrepância são as altas taxas de alarmes falsos e a complexidade de calibração dos algoritmos. Este artigo apresenta um algoritmo livre de calibração que pode ser aplicado em qualquer localidade. Os limites de decisão dinâmicos do algoritmo proposto são baseados nos dados históricos de tráfego, incorporando assim as variações típicas do fluxo ao longo do dia para reduzir os alarmes falsos. O modelo proposto obteve melhores resultados do que os algoritmos encontrados atualmente na literatura. # " Numerous automatic incident detection (AID) algorithms have been proposed in the literature over the last forty years. A myriad of algorithms of varied complexity, data requirements, and efficiency have been published in the literature. However, for desired levels of detection rate (DR), those algorithms yield unacceptably high false alarm rates (FARs) when implemented in the real world. In addition, AID studies verified with real data have been primarily based on computationally sophisticated methods whose extensive calibration and training efforts may discourage wide deployment by traffic management center (TMC) personnel. The fact that these models are typically configured to perform under very specific operational conditions for which they were calibrated makes their implementation not only difficult, but also inefficient when the operational condition drifts from the assumed norm. These problems have kept AID algorithms from being widely implemented, as it was found by a nationwide survey conducted in the United States involving 32 TMC (William and Guin, 2007), where it was concluded that only 12.5% of the centers claimed to have been using a fully functional AID algorithm. Another major problem of existing freeway AID models is universality (or transferability) (Abdulhai and Ritchie, 1999), which is the model’s ability to perform satisfactorily at different traffic scenarios/conditions with little or no recalibration efforts. The vast majority of the AID algorithms found in the literature are based on static (fixed) thresholds values for incident declaration, which leads to poor performance, as traffic state is mostly dynamic and fluctuates substantially throughout the day (Han, May, 1990). In addition, the calibration of some of the simplest detection algorithms relies on the availability of an incident dataset, whose development may be very time+consuming, especially considering that the necessary incident information recorded in crash reports, such as start+time and location, are usually inaccurate for AID algorithm calibration purposes. Therefore, these pieces of information would have to be corrected through cumbersome investigation of the traffic loop+detector data, or by the offline application of another incident detection algorithm. Traffic databases contain valuable information that can help traffic engineers discern normal and abnormal flow conditions, which is the primary objective of AID algorithms. The main research question addressed in this paper then becomes: “Could that information from historical data be used to develop a relatively simple model that dispenses calibration?” Therefore, the goal of this research was to develop and test an AID algorithm that eliminates the need of algorithm training without compromising performance. It combines historical (previous days) and current (real+time) data of traffic to improve universality and detection capabilities – acceptable levels of FAR, DR and mean+time+to+detection (MTTD). Such approach addresses the shortcomings of algorithms with fixed thresholds values by implementing demand+sensitive thresholds, thus enhancing the algorithm’s desirable transferability. $ " % $ # &' ' ' () The vast majority of the existing AID models are based on data gathered from inductive loop+ detectors. These algorithms can be basically grouped into four main categories: comparative, traffic+model+based, statistical, artificial+intelligence+based, and mixed models(Browne et al., 2005). They are briefly described below. Comparative algorithms are based on the comparison of traffic parameters (mainly occupancy) between adjacent stations, as downstream occupancy tends to drop while occupancy upstream of an incident is supposed to increase. The most traditional comparative model is the California and its derivatives (Payne, Tignor, 1978). The California algorithm checks the values of three variables that are based on the difference of occupancy between adjacent stations (occdf) against three predetermined thresholds (Thr1, Thr2, and Thr3). If all thresholds are exceeded in a particular moment, an incident is declared. Traffic model based algorithms are primarily based on traffic flow fundamentals, such as the McMaster algorithm proposed by Persuad and Hall (1989), which basically consists of defining a boundary between congested and uncongested flow+occupancy regions, and of identifying a speed threshold to distinguish congested from uncongested speeds. Other traffic+ model+based AID algorithms can be found in (Jin and Ran, 2009; Kuehne, 1980; Willsky et al, 1980). Statistical algorithms perform short+term prediction of traffic variables. If the predicted value deviates enough from the observed value, than an incident alarm is triggered. One of the earliest approaches is called the Standard Normal Deviate (SND) model (Dudek et al, 1974), in which the standardized value of a traffic control variable is checked against control limits that are based on the mean and the standard deviation of the data. The classical SND model formulation is SND = ( xˆ (t ) − x (t )) / S , where x(t) is the observed value and of the traffic variable, x(t ) is its predicted value (e.g.: mean), S is its standard deviation. If the variable SND exceeds a predetermined threshold, an incident alarm is triggered. Other statistical algorithms have used alternative values to the mean for x(t), including nonparametric regression (Tang and Gao, 2005). Other time series models used to forecast traffic volume for incident detection purposes have been tested (Ahmed and Cook, 1982; Cook and Cleveland, 1974). It is worth noting that no single forecasting technique would perform very well due to the high levels of noise inherent to 30+sec traffic data. Therefore, time+series models are usually combined with filtering techniques or other models to enhance prediction capability (Stephanedes and Chassiakos, 1993). Artificial intelligence based algorithms are models based on artificial neural networks (NNet) and fuzzy logic (Adeli and Samant, 2000; Cheu et al., 2004; Ishak and Al+Deek, 1999; Srinivasan et al., 2004). These methods have also been combined with other advanced models such as Wavelet theory (Ghosh+Dastidar and Adeli, 2003) and Wavelet transformation with linear discriminant analysis (Samant and Adeli, 2000). In this paper, two NNet models were applied for comparison: multi+layer feed+forward neural network (MLF), and wavelet+based multi+layer feed+forward (WMLF), which was based on the wavelet+filtering scheme proposed by Ghosh and Adeli (2003). Mixed models are those methods that combine different types of approaches. One of the most well+known mixed model is the Minnesota algorithm, which is a combination of statistical (time series filtering) and comparative types of algorithms (Stephanedes and Chassiakos, 1993). The Minnesota algorithm applies a low+pass filter (moving average) on the spatial differences in occupancies (occdf) before (yb) and after (ya) a particular time period. A normalized value of ya is checked against a threshold Thr1. If Thr1 is exceeded, a normalized value of yb + ya is compared with a second threshold Thr2. If Thr2 is also exceeded, an alarm is triggered. In this paper, the existing algorithms used for comparison were California (comparative), Minnesota (mixed), SND (statistical), and two NNet models (Artificial+ intelligence+based). $$ ( * + ) , ) As it usually occurs in detection systems, there are trade+offs to be considered among DR, FAR, and MTTD. In the case of freeway AID, for desired levels of DR, FARs have been unacceptably high for operational purposes (William and Guin, 2007). The high occurrence of false alarms in freeway AID systems can be attributed to several factors. First, there are situations where traffic may exhibit incident+like patterns when in fact there is no incident, such as in the presence of freeway bottlenecks. Figure 1 illustrates historical 5+min occupancies between adjacent stations on I+880N; US means upstream station and DS means downstream station. The occupancy profiles represent the median of 5.5 months of data. ! "# " $ ! %& '& ' #: Historical 5+min occupancies for two neighboring stations on I+880N Other common situations where substantial differences in occupancy between neighboring VDSs may occur are: significant on+ and off+ramp traffic volumes; compression waves (Payne and Tignor, 1978); difference in operational speed; space between stations; and specially, the high noise of 30+second data. Due to a number of reasons (e.g.: presence of trucks), traffic may assume various states within 30 seconds. This has been widely recognized as one of the main challenges faced by AID algorithms (Abdulhai and Ritchie, 1998; Stephanedes and Chassiakos, 1993). All of the above mentioned factors contributing to false alarms can be mitigated if historical information of traffic is incorporated into the AID functionality. In addition to these factors inherent to traffic, AID models proposed in the literature present two major problems that are conducive to increasing levels of false alarms, namely calibration complexity and lack of universality. As for the former, even the simpler algorithms require considerable calibration efforts to determine the best algorithm threshold values for each individual, or pair of, stations. This extensive calibration complexity may lead to lack of universality, which leads to poor performance (high FAR). More detailed information on AID models transferability may be found in (Mak and Fan, 2005; Stephanedes and Hourdakis, 1996). The aforementioned shortcomings have kept AID algorithms from being widely implemented in the United States. This was the conclusion of a recent North American survey on the use and conception of AID algorithms at 32 TMCs located throughout the US and one TMC in Ontario, Canada (William and Guin, 2007). The following are some important findings revealed by the survey, which was responded by key managers from the TMCs. • 70% of the respondents considered the existing methods of incident detection to be inefficient. • Even though 53% of the centers have an AID algorithm integrated to their system, only 12.5% considered their AID to be operational. • The main reported reasons for the limited use of AID algorithms were, in order: 1) high rate of false alarms; 2) difficulty in calibration; 3) low detection rates. * * ** Three key characteristics of the presented method should to be highlighted in advance: 1) the algorithm is very simple and requires no training; 2) it is self+learning, as it needs no human intervention and becomes more efficient with time; and 3) its detection is based on a dynamic traffic+demand+sensitive threshold. The fundamental idea of the proposed mixed algorithm is to identify what are the likely values of the 30+sec occupancy differences between neighboring VDSs for a particular 5+min period of the day. The set of likely values is based on historical 5+min occupancy differences observed in previous days. 30 , the ith difference of 30+sec occupancy between two adjacent Consider stations inside the 5+min period (j), for a day (d) of the week. Notice that i=1,2,…10, as there are ten 30+sec observations inside a 5+min period. Assuming that for a particular (j,d) (1) occdf 30 sec ( i ) ( j , d ) ~ N ( 30 sec , σ 2 30 sec ) This means that, for a (j,d) pair, the historical 30+sec differences in occupancy between two adjacent stations are normally distributed—this assumption was tested and confirmed, as it will be shown later in this paper. If )30sec and σ230sec are estimated, then a one+sided region of probable values of 30 , can be constructed. Current observations falling outside that interval will be considered incident. The parameters )30sec and σ230sec can be estimated based on the historical 5+min loop detector occupancies, as follows. Let 5 , be the difference of 5+min occupancies between two adjacent stations for a particular 5+min period j of the day d of the week. Let )5min and σ25min be the historical mean and the variance of 5 , , of all for a particular (j,d) pair. Since )30sec = )5min ( 5 , is the mean of 30 , , )30sec can be simply estimated by 5 min that is, by X 5 min . 5 , . The parameter σ230sec can also be estimated from the historical data of Assuming that 30 , observations for a particular (j,d) are independent, then 5 , is the mean of 30 , . Therefore, σ25min= σ230sec/10 , because 2 )30sec and σ 30sec, the only parameters to be estimated in the proposed AID model, can be easily estimated from the historical values of the 5+min difference in occupancies. Equation (1) becomes: (2) occdf30sec (i)(j,d)~N() 5min ,10σ 2 5min ) For a desired level of FAR (α), a one+sided interval that comprises (1+ α)*100% of the 30 can be defined, with its upper+limit value becoming the threshold of the model for the particular 5+min period. (3) Thr = NormInv(X 5min ,10 S 2 5min ,1 − α ) where NormInv is the inverse of the Normal cumulative distribution function. If an 30 , exceeds Thr, an incident alarm is triggered. The threshold continuously changes every 5 minutes, accounting for changes in traffic based on its typical behavior. . % " .# / '0 ' 1 2 2 /'3 + The data used in this research came from the Freeway Performance Measurement System (PeMS, 2009). A total of 40 lane+blocking incidents were collected. The process of mining incidents is described as follows. First, the California Highway Patrol (CHP) incident logs were studied. Then, for a reported incident, the corresponding traffic data for the VDSs surrounding the stated location were scrutinized. The start+time reported on the CHP report was checked against the time when traffic flow was first disturbed; as expected, they usually did not match on the minute+level. Hence, in this research, the start+time of an incident was the apparent start time, defined as the time interval immediately before the traffic disturbance was first observed, an approach that is not ideal but that has been implemented by other AID studies (Mak and Fan, 2006). This incident data collection methodology was used to collect 40 incidents, from which 20 were randomly selected for training, and the remainder for testing. It is important to note that the training data set was used only by the traditional algorithms, as the proposed AID algorithm requires no calibration. All incidents were collected from June 1 through November 15 of 2006 (5.5 months) on the northbound facility of interstate 880 (I+880N), a freeway that had been studied before and was deemed to have one of the highest crash frequencies in the San Francisco Bay Area, California (Skabardonis et al., 1997). The I+880N freeway is a 74+km facility, of which 55 miles are served by a high+occupancy+vehicle (HOV) lane. It has a total of 311 loop detectors that form 75 VDSs. Its Annual Average Daily Traffic (AADT) is approximately 125,000 veh/day (north+bound facility only). Traffic data free of incidents were culled to evaluate the FAR of the models. Data along the freeway where relatively large recurrent differences in occupancy between neighboring stations occurred were identified and selected, as they are conducive to the occurrence of false alarms. Besides the recurrence of the differences in occupancy, the absence of incident records on the CHP logs was also considered before the data were deemed incident+free. Seventeen incident+free cases were selected, of which 9 were randomly selected for training and 8 for testing, resulting in 46.5 hours of traffic data. . $ 00 '/ ' , ( 3 In this paper, five existing AID algorithms were evaluated for comparison purposes, namely California (comparative), Minnesota (mixed), SND (statistical), MLF, and WML (artificial intelligence). The comparison was based on performance curves of DR x FAR, which have been widely used in previous AID studies (Browne et al., 2005; Petty et al., 2002; Teng and Qi, 2003). The MTTD was also considered in the evaluation. In the application of the California Algorithm, each threshold (Thr1, Thr2, and Thr3) was tested from 0.05 through 1.00, with increments of 0.05 on the training dataset, which resulted in a total of 8,000 (203) combinations of thresholds. For each level of DR (from 0.8 to 1.0), the model with the minimum FAR was selected. If two models yielded the same FAR, the one with the lowest MTTD was chosen. The selected models were then evaluated on the testing dataset. This model selection criterion was applied in all 5 existing algorithms. For Minnesota algorithm, each threshold (Thr1 and Thr2) was tested from 0.05 to 1.0, with 0.05 increments, resulting in a total of 400 set of parameters tested. Window sizes of yat and ybt were 10 (5 min) and 6 (3 min) observations, respectively, the values suggested by the authors who introduced the model (Stephanedes and Chassiakos, 1993). The best models were then evaluated on the testing dataset. In the application of the SND algorithm, besides threshold Thr1, the window size (WS) of the look+back interval was also tested for different values—4, 6, 8, and 10 minutes. Since Thr1 ranged from 0.5 to 1.5 with 0.1 increments, a total of 584 (146x4) models were tested. The WS=8min achieved the best training performance, therefore it was the WS size used in the testing stage. The MLF algorithm evaluated consisted of eight neurons (upstream and downstream occupancy at times t 3, t 2, t 1 and t) in the input layer, ten neurons in the single hidden layer and one output neuron. As activation functions in the hidden and output layers, tangent sigmoid and linear transfer functions were used, respectively. As for the WMLF, it had wavelet coefficients as neurons in the input layers. The architecture of WMLF algorithms consisted of six neurons in the input layer, a single hidden layer with ten neurons, and one output neuron. As it was the case for the MLF, tangent sigmoid and linear transfer functions were used as activation functions in the hidden and output layers. As previously stated, this algorithm was based on the filtering scheme proposed by Ghosh and Adeli (2003). In the application of the proposed algorithm, the first step is to select the historical data that will be used by the algorithm. For each 5+minute period, the algorithm threshold changes according to the 5+minute occupancy differences observed in the previous days, as indicated by Equation 2. Since traffic is known to vary by day+of+the+week, only days with expected similar traffic behavior should be considered in the historical sample. In this paper, the following groups were considered to be homogeneous: Mondays (business days); Tuesdays, Wednesdays, Thursdays (business days); Fridays (business days); and Saturdays, Sundays, and non+business days. Therefore, when applying the algorithm on a Wednesday, all previous Tuesdays, Wednesdays and Thursdays that are business days are considered in the historical sample. Since the threshold depends on the variance of occdf5min(j,d) computed over previous days, it is important that outliers be identified and removed. In this work, occdf5min(j,d) observations lying outside the interval ˆ5 min ± 2σˆ 52 min were discarded from the historical sample. The threshold for the 5+minute period of interest is determined from Equation (3). Initially, α=0.01% was chosen. It is important to note that the value of α represents the desired false alarm rate, not a parameter to be calibrated. Figure 2 shows the time+varying threshold obtained from the proposed AID model. This figure shows occdf 30sec for accident number 31 (Acc31). The vertical dotted lines specify the apparent start+ and end+times of the accident. The relatively flat blue line is ˆ5 min , the historical sample average of occdf 5 min( j , d ) . As shown in the same figure, the threshold stays high during the PM peak+period (until 07:00pm) and decreases afterwards, allowing the algorithm to detect the accident from its start. Figure 2b shows an incident+free case. Notice that the algorithm’s thresholds are high enough to thus avoiding excessive false alarms, indicating that the observed differences of occupancy between the stations are actually recurrent. Hence, by considering historical information, the proposed AID algorithm avoided false alarms around 07:15am, although it sounded one around 06:50am, which have been promptly deemed to be false if a persistence test of 1 observation (30+sec) was used. # +,' $ + * # ) +,' $ + / " ) * " * $ $ ( ) () * $ ( ) - . ! (* $ ( ) () ( ) (* * $ (a) ' ( ) - ( ( ) ( . ! (b) $: a) Accident #31 detected by the proposed algorithm. b) Incident+free #02. .- 4 ' Figure 3 shows the comparison of the existing models and that of the proposed model for three different values of α. The results show that the proposed algorithm presented DR=0.95 for all α levels, with the lowest FAR. On that level, the second best algorithm was the SND, with FAR twice as high. For α=0.1%, the relatively low FAR (0.25%) is achieved on the expense of a high MTTD (4.3 minutes), which is an expected result, since a lower α means higher thresholds.. A lower MTTD is obtained by increasing α, which consequently increases FAR, as the thresholds are lowered. For an alpha of 0.5%, FAR increased to around 0.6%, and MTTD is 2 minutes. The proposed algorithm assumes that, for a particular 5+min period of the day, the historical values of occdf30sec are independently and normally distributed. To verify the validity of this assumption, 30+sec occupancy data of a VDS (#400983) for six days were collected (October 4, 5, 10, 11, 24, and 31 of 2006). Chi+square goodness+of+fit tests were applied within each of 288 five+minute periods of the day. For a significance level of 5%, normality tests did not rejected the null hypothesis of normality in any of the 288 samples. Since occdf30sec(i)(j,d) is the difference between adjacent stations, it can be concluded that it also can be considered normally distributed. 5 " " When comparing AID algorithms, one should consider not only DR, FAR, and MTTD, but also two foremost features of AID models: ease of implementation, and universality, which are widely recognized as the most critical problems encountered in existing algorithms (Abdulhai, B., and Ritchie, S., 1999). ! 1# α6 α6 ) α6 )* '0 3 - * * ! ' 2 " *) * ! ** &3' ) # ) 4/ ) 5 4 * ) /+0 ' -6 Detection performance of the proposed and existing model 5# , )0 ) ' The proposed AID algorithm requires no training, that is, no calibration. This certainly encourages implementation as calibration may require significant time and human efforts that are not always available. In addition, the calibration of existing approaches also requires the availability of an incident dataset, which must contain relatively accurate information such as start+time and location. As aforementioned, start+times reported in incident logs are not accurate enough for AID algorithm calibration purposes. Therefore, the traffic data must be scrutinized so the apparent start+time can be determined. Such process may be very time+consuming. Even considering that a well documented incident database is available, and that the parameter calibration is fairly simple to conduct, the calibration process of existing models will single out a parameter—or a set of parameters—that may perform well only in a narrow scope of traffic situations. In this case the model lacks transferability, which is discussed in the next section. Another positive feature of the proposed model that considerably simplifies implementation is that it is self+learning, that is, it improves its detection performance by itself, with no human intervention. The more traffic data are received by the TMC center, the greater the ability of the model to capture the typical behavior of traffic. 5 $ " '4 ' 7 , +' ' 8 In addition to its ease of implementation, which is a highly desirable attribute towards model universality, the presented model is transferable to any traffic situation because it is based solely on the typical behavior of traffic for the particular time and location. Therefore, the logic of the model can be applied regardless of the type of roadway geometry, functional classification, and very importantly, time. In the proposed model, the recurrent differences in occupancy are taken into consideration in the computation of the dynamic thresholds. It is worth noting that some authors state that an AID algorithm is transferable if it works successfully in different freeway facilities. This is not necessarily true, as the VDSs tested on two or more freeways may have similar traffic characteristics. Even though the proposed AID was tested in a single freeway facility, the I+880N presents a variety of geometry characteristics along its 73.6 km. An example of the universality of the proposed model is shown in Figure 4. In this case (Acc28), the differences in occupancy caused by the accident were not large enough to make the existing models detect the incident; even the California model that yielded the highest level of DR was not able to detect it. Since during this particular time of the day the differences in occupancy are historically low—as indicated by the low thresholds themselves—the thresholds of the proposed model were low enough to detect the incident. # +,' $ + " $ $ (* () ( ( * $ ' ( - ( (* () ( . ! .6 Incident detected by the proposed model. 5$ ) ,'3 / 3'/ ' Even though the presented method provides thresholds for decision+making based on a specified desired FAR α, TMC operators can alternatively look at the probability of observing the observed occdf(i)30sec value or higher, which represents a “p value”. Therefore, TMC users may initially set a high α—say, 2 %—and check the p value associated with the alarm before taking further action. If the alarm is triggered by an observation that presents a very low p value, the TMC operators may consider it to be an incident. They can also wait for the next observation and check its p value before making the decision. Another advantage of having a probability associated with the occdf(i)30sec observation is that p values may indicate the severity of the incident, as extremely low p values mean that large occdf30sec are observed. Therefore, the TMS personnel may want to direct the response efforts by prioritizing those incidents where the p values are lowest. 5/' * , ) / Figure 3 showed that the proposed approach provided the best detection performance. For DR equal to 95%, Minnesota and SND algorithms presented higher FAR. When α=1% the proposed model MTTD is 0.5 minutes higher than that of the SND algorithm, but the FAR of the proposed method is half of that of SND. The comparatively low FARs are attributed to the proposed algorithm’s accountability for recurrent differences in occupancy. It is worth noting that even if the proposed and existing AID models had performed equally, the proposed algorithm would still be valuable as it is much simpler to implement as well as more universal. 9 " : ! In addition to the better performance provided by the proposed model based on the traditional measures of effectiveness—DR, FAR, and MTTD—the presented AID model has other significant advantages when compared with existing models. The most important ones are: • Ease of implementation. The presented algorithm dispenses calibration .The method uses only historical average 5+min and current 30+sec occupancy data, which are commonly available in most TMS databases. Another feature that simplifies implementation significantly is that the algorithm is self learning, that is, it requires no human intervention. The more traffic data are inputted into the model, the greater its ability to capture the typical behavior of traffic, thus automatically improving incident detection performance. • Universality. The proposed algorithm is designed to work in any spatio+temporal traffic configuration, as its dynamic threshold is based on the normal behavior of traffic on a particular pair of neighboring stations at a specific 5+min period of the day. Therefore, no adjustment needs to be performed when the model is applied at different sites and in different times of the day. As recommendations for future work, other techniques of outlier removal may be investigated, as outliers in the historical 5+min occupancy data may negatively affect the performance of the model. The ideal outlier removal technique would eliminate all high occdf5min observations coming from special events and erratic data while keeping the observations coming from normal operations. Also, the proposed approach can be tested using different traffic parameters, such as the 30sec difference of speeds, and information based on traffic flow fundamentals could be incorporated to enhance performance. Other recommendations include the assessment of using persistence tests, and the evaluation of other threshold aggregation levels (other than 5+minutes). /; 2 3 ) This research was entirely supported the Federal Highway Administration through its Dwight Eisenhower Transportation Graduate Fellowship Program. The authors also would like to acknowledge the Freeway Performance Measurement System (PeMS) for the data and all the richness of information provided by the system Abdulhai, B., and S. Ritchie (1999). Enhancing the Universality and Transferability of Freeway Incident Detection using a Bayesian+Based Neural Network. Transportation Research Part C, 7, 261+280. Adeli, H., and A. Samant (2000). An Adaptive Conjugate Gradient Neural Network+Wavelet Model for Traffic Incident Detection. Computer aided Civil and Infrastructure Engineering, 15, 251+260. Ahmed, S., and A. Cook (1982). Discrete Dynamic Models for Freeway Incident Detection Systems. Transportation Planning and Technology, 7, 231+242. Browne, R., S. Foo, S. Huynh, B. Abdulhai and F. Hall (2005). Comparison and Analysis Tool for Automatic Incident Detection. Transportation Research Record, 1925, 58+65. Cheu, R., D. Srinivasan, and W. Loo (2004). Training Neural Networks to Detect Freeway Incidents by Using Particle Swarm Optimization. Transportation Research Record, 1867, 11+18. Cook, A., and D. Cleveland (1974). Detection of Freeway Capacity+Reducing Incidents by Traffic+Stream Measurements. Transportation Research Record, 495, 1+11. Dudek, C., C. Messer, and N. Nuckles (1974). Incident Detection on Urban Freeways. Transportation Research Record, 495, 12+24. Ghosh+Dastidar, S., and H. Adeli (2003). Wavelet+Clustering+Neural Network Model for Freeway Incident Detection. Computer aided and Civil Infrastructure Engineering, 18, 325+338. Han, L., and A. May (1990). Traffic Flow Characteristics of Signalized Arterials under Disturbance Situations (Research Report UCB ITS RR 90 12A). Berkeley, CA: University of California Berkeley Institute of Transportation Studies. Ishak,, S. and H. Al+Deek (1999). Performance of Automatic ANN+based Incident Detection on Freeways. Journal of Transportation Engineering, 125 (4), 281+290. Jin, J. and B. Ran (2009). Automatic Freeway Incident Detection Based on Fundamental Diagrams of Traffic Flow. Transportation Research Record, 2099, 65+79. Kuehne, R. (1989). Freeway Control and Incident Detection using a Stochastic Continuum Theory of Traffic Flow. Proceedings of the 1st International Conference on Application of Advanced Technology in Transportation Engineering (pp. 287+292). New York: ASCE. Mak, C. and H. Fan (2005). Transferability of Express Incident Detection Algorithms to Singapore and Melbourne. Journal of Transportation Engineering, 131 (2), 101+111. Mak, C. and H. Fan (2006). Algorithm fusion for detecting incidents on Singapore's central expressway. Journal of Transportation Engineering, 132(4), 321+330. Payne, H. and S. Tignor (1978). Freeway Incident+Detection Algorithms Based on Decision Trees with States. Transportation Research Record, 682, 30+37. PeMS. (2009). Freeway Performance Measurement System, Version 10.1. Retrieved July 1, 2009, from http://pems.eecs.berkeley.edu Persuad, B. and F. Hall (1989). Catastrophe Theory and Patterns in 30+Second Freeway Traffic Data – Implications for Incident Detection. Transportation Research – Part A, 23A (2), 103+113. Petty, K., M. Ostland, J. Kwon, J. Rice, and P. Bickel (2002). A New Methodology for Evaluating Incident Detection Algorithms. Transportation Research Part C, 10, 189+204. Samant A. and H. Adeli (2000). Feature Extraction for Traffic Incident Detection Using Wavelet Transform and Linear Discriminant Analysis. Computer aided and Infrastructure Engineering, 15, 241+250. Skabardonis, A., K. Petty, R. Bertini, P. Varaiya, H. Noemi, and D. Rydzewski (1997). I+880 Field Experiment: Analysis of Incident Data. Transportation Research Record, 1603, 72+79. Srinivasan, D. X. Jin, and R. Cheu (2004). Evaluation of Adaptive Neural Network Models for Freeway Incident Detection. IEEE transactions on Intelligent Transportation Systems, 5 (1), 1+11. Stephanedes, Y. and A. Chassiakos (1993). Application of Filtering Techniques for Incident Detection. Journal of Transportation Engineering, 119 (1), 13+26. Stephanedes, Y. and J. Hourdakis (1996). Transferability of Freeway Incident Detection Algorithms. Transportation Research Record, 1554, 184+195. Tang, S. and H. Gao (2005). Traffic+Incident Detection+Algorithm Based on Nonparametric Regression. IEEE Transactions on Intelligent Transportation Systems, 6 (1), 38+42. Teng, H. and Y. Qi (2003). Detection+Delay+Based Freeway Incident Detection Algorithms. Transportation Research Part C, 11, 265+287. William, B. and A. Guin (2007). Traffic Management Center Use of Incident Detection Algorithms: Findings of a Nationwide Survey. IEEE Transactions on Intelligent Transportation Systems, 8 (2), 351+358. Willsky, A., E. Chow, S. Gershwin, C. Greene, P. Houpt, and A. Kurkjian (1980). Dynamic Model+Based Techniques for the Detection of Incidents on Freeways. IEEE Transactions on Automatic Control, AC 25(3), 347+360. 3 < 3 (manoel@det.ufc.br) Departamento de Engenharia de Transportes, Universidade Federal do Ceará. Campus do PICI, s/n – Bloco 703 – CEP. 60455+760 – Fortaleza, CE, Brasil. (lhan@utk.edu) Department of Civil and Environmental Engineering, University of Tennessee Knoxville, 223 Perkins Hall, 37996, TN, EUA (ysjeong@eden.rutgers.edu) ! (mjeong@rci.rutgers.edu) Department of Industrial and Systems Engineering, Rutgers University, 640 Bartholomew Road, 08854, NJ, EUA