IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION …cvrr.ucsd.edu/publications/2012/Morris_ITS2012.pdf · vehicle proﬁles, the differences between different vehicle types and instantaneous

IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS - TO APPEAR 2012 1

Real-Time Video-Based Traffic Measurement andVisualization System for Energy/Emissions

Brendan Tran Morris, Member, IEEE, Cuong Tran, Student Member, IEEE, George Scora,Mohan Manubhai Trivedi, Fellow, IEEE and Matthew J Barth, Senior Member, IEEE

Abstract—It has become increasingly important to monitor thestate of roadways in order to better manage traffic congestion.Sophisticated traffic management systems are in development toprocess both the static and mobile sensor data that provide trafficinformation for the roadway network. In addition to typicaltraffic data such as flow, density, and average traffic speed,there is now strong interest in environmental factors such asgreenhouse gases, pollutant emissions, and fuel consumption fromtraffic. It is now possible to combine high resolution real-timetraffic data with instantaneous emission models to estimate theseenvironmental measures in real-time. In this paper, a systemis described that estimates average traffic fuel economy, CO2,CO, HC, and NOx emissions using a computer vision-basedmethodology in combination with vehicle specific power basedenergy and emission models. The CalSentry system provides notonly the typical traffic measures, but also gives individual vehicletrajectories (instantaneous dynamics) and recognizes vehiclecategories which are used in the emission models to predict envi-ronmental parameters. This estimation process provides far moredynamic and accurate environmental information compared tostatic emission inventory estimation models.

I. INTRODUCTION

W ITH increased roadway congestion, it has becomecritical to monitor the state of the roadway network

through a variety of means. Over the last decade, there hasbeen a tremendous amount of research in intelligent trans-portation systems (ITS) for advanced traffic monitoring andmanagement. Traffic management centers (TMC) around theworld are becoming increasingly sophisticated. They bringin data from large networks of sensors for analysis in orderto better manage overall traffic. Efficient operation of thesecenters requires both an up-to-date view of current conditions,for speedy response, as well as historical data for modeling,planning, and prediction.

A prime example of such a system is California’s Per-formance Measurement System (PeMS) [1] which gathersmeasurements from 30 thousand inductive loops embeddedin the highway and distributed across the state in addition topolice incident reports and lane closure information. PeMS hasbeen a great success because it has made the fundamental flow,occupancy, and speed data along with basic traffic calculations

B. Morris is an Assistant Professor at the University of Nevada, Las Vegas,NV, 89154-4026 USA e-mail: [email protected]

C. Tran and M. Trivedi are members of the Computer Vision and RoboticsResearch Laboratory, University of California, San Diego, CA, 92093-0434USA e-mail: [email protected], [email protected].

G. Scora and M. Barth are members of the Center for EnvironmentalResearch and Technology, University of California, Riverside, CA, 92507USA email: [email protected], [email protected].

Manuscript received Sept. 23, 2011.

accessible to the research community, thus inspiring novel newtraffic management outcomes.

More recently, video cameras have become popular inTMCs for human monitoring and verification to augment ITSdata elements such as loops. Cameras provide complementaryanalysis difficult to manage using traditional sensors. Vision-based systems have been developed in the past decade toallow real-time traffic flow, vehicle classification, tracking, andtrajectory analysis [2]–[5]. Cameras have also been integratedin multi-modal frameworks for structural health monitoringand event detection [6].

While capacity and congestion have historically been themajor motivating factors of transportation management, newperformance metrics have recently garnered attention. In ad-dition to standard traffic metrics, there is a strong interest intraffic-related emissions now in terms of

1) pollutants (e.g. carbon monoxide (CO), hydrocarbons(HC), nitrides of oxygen (NOx), and particulate matter)

2) greenhouse gases (e.g. carbon dioxide (CO2))3) and energy (fuel consumption).Estimating the emissions inventory for vehicles traveling

on the roadway network is an active field due to emissionrequirement from government institutions such as the U.S.Environmental Protection Agency (EPA) and the CaliforniaAir Resources Board (CARB). Both the EPA and CARB havesophisticated emission models [7], [8] that can be used todetermine emissions for specific scenarios and most roadwayplanning must utilize these models to determine the impacts offuture activity. The transportation community is now beginningto see the value of combining both real-time transporta-tion data and emissions modeling to predict instantaneousemissions or energy usage on a road network on a link-by-link basis. Unfortunately, there is no PeMS-like systemfor accurate roadway emissions measurements. Attempts havebeen made to utilize link-based traffic volumes and averagespeeds along with speed-emissions curves for estimation [9]but these approaches lack sensitivity. They do not account forvehicle profiles, the differences between different vehicle typesand instantaneous activity, which drastically affect emissions.

In order to provide real-time link-based emissions (andfuel economy), this work has developed the CalSentry systemwhich combines the VEhicle Classifier and Traffic flOWanalyzeR (VECTOR) [10] system, a computer vision-basedhighway measurement system, with vehicle specific power(VSP) [11] based energy/emission profiles derived from thethe Comprehensive Modal Emission Model CMEM [12] andthe MOtor Vehicle Emissions Simulator (MOVES) [7]. This


system provides subtle vehicle dynamics (instantaneous speedand acceleration) through visual tracking along with cate-gorization of the type of vehicles on the road in order toaccurately estimate vehicle-specific emissions. The systemcould be deployed within a larger sensor network to provide astreaming data source for traffic management system such asPeMS. Thus providing useful real-time information for policymakers, planners, and health officials and facilitating furthertransportation related emissions research.

II. RELATED STUDIES

The world’s rapidly growing and expanding populationhas led to traffic congestion even in the best planned roadnetworks. This congestion results in a loss of time and produc-tivity and contributes a high economical cost. In order to tacklethe congestion problem, without continual road constructionprojects, traffic management and control approaches have beenadopted to better utilize the existing roadway infrastructure.However, in order to develop effective management or controlstrategies, data is needed. Historical data is needed to learnand develop models while real-time measurements can provideup-to-date indicators of performance for prompt response. Inaddition, this data is needed over large coverage areas withvaried conditions. Effective monitoring systems must thereforebe scalable, provide distributed and cooperative sensing, berobust to a wide range of environmental conditions, and haveefficient transmission and storage.

The dominant sensor for traffic management has been theinductive loop sensor which is able to detect the presence ofa vehicle based on an induced magnetic field. This simplespot detector provides the count of vehicles that have passedover it (flow) as well as the amount of activation (occupancy)in a time period. By collecting the loop readings from manysensors, traffic researchers have built complex models foranalysis. PeMS [1] which was developed at UC Berkeley (andis now in Caltran’s control) provides raw loop readings as wellas fundamental traffic performance measures. PeMS collectsand stores the data from over 34,000 inductive loops onCalifornia’s highways making it indispensable for researchers.

However, inductive loops are limited because they are costlyto install and maintain (only 62% of California’s PeMS loopsensors are in working order), making the search for alternativesensing solutions [?] or augmentation schemes [?] appealing. Video cameras have emerged as a popular ITS devicebecause of large spatial coverage or field of view (FOV)which allows capture of higher order dynamics in vehiclesand traffic, rich information content for more complex analysis(e.g. classification), and many TMC have already installedthem for human observation.

A. Visual Traffic Monitoring

Highway monitoring has been one of the oldest applicationsfor vision researchers. The inherent structure of roads coupledwith a vehicle’s rigid body constrains the vision processes.Early vision-based traffic monitoring researchers looked tomimic the popular inductive loop sensor counts by manuallydefining virtual loops in the camera FOV [13]. While easy to

manage and effective, the virtual loops did not take advantageof the spatial coverage afforded by the camera, reducing thewide FOV into several small point sensors. Subsequently, mostresearchers began to focus on moving object detection andtracking [14]. In this paradigm, a count is generated for everytracked vehicle [10], [15].

The spatial sensing and wide area coverage which makesvideo an attractive monitoring sensor also provides the greatestdifficulty. In order to properly count vehicles, each and everyvehicle must be detected with a single sensor. Camera place-ment and view are critical for successful deployment. Imagingprovides roadway coverage over long distances but also causesperspective distortion which greatly affects the apparent sizeof vehicles and leads to occlusion. Significant effort has goneinto developing detection methods which can resolve occlusionsuch as feature grouping in the image plane [15] or in 3D spaceusing multiple homography transformations [16].

Cameras also have difficulties dealing with changing envi-ronmental conditions. The sun’s normal course during the daycauses challenging illumination conditions due to cast shad-ows. Researchers have attempted to distinguish cast shadowson the roadway from vehicles using shadow detection andsuppression techniques [17]. In addition to shadows caused bylighting, it is quite difficult to operate vision systems at night,reducing the effective operating time. Major efforts have begunto develop techniques to detect and track vehicle headlights[18] to avoid use costly low light sensitive or infrared cameras.

B. Vehicle ClassificationThe switch to video-based traffic monitoring is particular

useful for vehicle classification because of the appearanceinformation contained in an image. Loop-like sensors onlygenerate a one dimensional signature which makes it difficultto resolve differences between vehicles (large and fast movingvs. small and slow). Generally, these systems count the numberof axels to only distinguish between large and small vehicles.

The review by Buch et al. [19] devotes two large sectionsto top-down (object-based) and bottom-up (part-based) visualclassification techniques for urban traffic. It is noted thatoften times classification degenerates into a detection problembecause the techniques are designed for matching. In fact, evena recent vehicle classification paper only makes a distinctionbetween two different classes of vehicles based on stablefeatures [16]. Early work by Gupte et al. [20] classifiedvehicles by their length. However, length measurement pre-cision was found to be low and not flexible to camera views.More detailed classification has been tackled using shape andappearance techniques [10], using a linearity feature [21], orexplicit vehicle models. Edge matching techniques have beendesigned using 3D wire-frame vehicle models [22]. Thoughthe 4 models were quite simple and had low resolution,they only operated at 5Hz. More generic and adaptive 3Dmodels have been explored to provide a deformable vehiclemodel with higher resolution [23]. Results were shown upto a difficult 5 class problem where a distinction was madebetween 2 and 4 door sedans. Despite these efforts, it is stilldifficult to leverage high resolution imaging while maintainingthe computational efficiency required for real-time monitoring.


Vehicle Type

Parameters

Tracks

Traffic Statistics Roadway

Parameters

Emissions Parameters

Emission Statistics

Vehicle Detection Tracking

Classification Traffic Statistics

Emission Modeling

Visualization

Static Database Dynamic Database

Fig. 1: CalSentry block diagram: VECTOR subsystem in purple performs visual tracking, classification, and traffic statisticmeasurement, emission modeling module in green, visualization module in orange, all connected to static calibration databasesand dynamic databases to store highway performance measurements.

C. Emissions/Energy

Environmentalists and health officials have long been con-cerned with the effects of air pollution on air quality, butonly recently has there been a major shift in focus to thetransportation sector. Traditionally, monitors which measurepollutants in the air in parts per million and parts per billionare used to evaluate air quality. Unfortunately, the number ofmonitoring sites is limited and existing sites provide sparsespatial data that does not necessarily express traffic emissionsspecifically since contributions from other pollutant sourcesmay be included. Monitoring sites cannot accurately depictthe spatial distribution of pollutants or target area for fo-cused surveys, nor can they attribute emission contributions toindividual vehicles. Mobile measurement systems have beenempolyed and have shown that high levels of pollutants arefound near highways which significantly exceed maximumvalues reported by fixed measurement monitors [24].

In order to isolate the effects of transportation emissions,a wide range of modeling techniques have been adopted [9].The simplest models are easy to calculate and rely on averagespeed, but do not account for real world driving characteristics[25]. More complex modal models operate at a higher timeresolution (seconds) and account for more detailed vehicleand traffic characteristics such as specific engine operation andvehicle movements. The modal models have gained traction re-cently because supporting data can more easily be obtained viaGPS [26] and they can be integrated into microscopic trafficsimulation models [27]. While promising, these techniques aredifficult to scale to large areas. Communication networks mustbe established and there must be high penetration rate for GPS-based emission calculations. Microscopic traffic and emissionssimulation is computationally expensive for large networks

because trajectories and emissions must be calculated for eachvehicle at each simulation time step over the entire networkand are affected by errors in the traffic models themselves.

Currently, there is no tool available to calculate trans-portation related emissions in real-time for a large numberof vehicles. In addition, there is no way to present thisinformation to stakeholders in order to manage or plan futuredecisions.

III. CALSENTRY HIGHWAY EMISSION MANAGEMENTSYSTEM

This manuscript presents CalSentry, the first real-time inte-grated highway transportation measurement and managementsystem for emission/energy estimation. This vision-based sys-tem combines four major components as shown in Fig. 1:

1) visual traffic measurement,2) dynamics-based emission estimation,3) real-time visualization,4) and a database for record keeping.

The elements in purple comprise the parts of the VECTORhighway monitoring module [10] which is a visual trackingand analysis system suitable for distributed traffic understand-ing [5]. The emissions modeling and estimation block in greenutilizes VECTOR analysis in order to estimate the amounts ofpollutants produced by vehicles on the roadway using real-time emission modeling [7], [12]. Adhering to a frameworkdesigned for thematic contextualization [28], measurementsand model estimations are stored and utilized for appropriatevisualization of system output (orange block). As a testamentto its robustness, the CalSentry system has been in continuousdaily operation at a single site since early 2011, collecting


(a)

(b) (c)

Fig. 2: CalSentry visualization (a) VECTOR system for vehicle tracking and classification. (b) Real-time plots of vehicle countsand emissions. (c) Google map with highway color-coded based on emission measurement; updated in 30 second intervals.

highway data during the daylight hours (approximately 5:00 -19:00).

At the host site, a traffic operator can view live highwayvideo feeds; both in raw and processed form. Fig. 2a presentsthe output of the VECTOR system. The highway video isprocessed in real-time to display object detection and trackingresults. Tracked vehicles have a color-coded bounding box toindicate the current emission score (based on dynamics andvehicle type) with red indicating a higher score. In addition, onthe left of Fig. 2b in red and white are moving time-series plotsof of highway flow. The plots are updated every video frame togive the instantaneous count of vehicles traveling either northor south while providing a short 30 second history. Using theVSP approach [11] (described in Section V-B), the roadwayemissions are estimated based on tracking information andvehicle type (determined by visual processing). Similar tothe moving flow plots, the instantaneous VSP value with 30second history is displayed in yellow and blue. The time-seriesplots are intended to show the evolution of conditions on theroad. On the right of Fig. 2b are two bins representing the totalaccumulated emissions in the north and south directions in asliding 30 second window. The bins are color-coded, yellow,orange, and red, to indicated low, medium, and high amountsof greenhouse emissions.

The diagnostics plots of Fig. 2b provide immediate up-to-date measurements but are quite variable due to trafficcongestion conditions. The emission score, which includes thefour greenhouse gases and pollutants {CO2, CO, HC, NOx},is accumulated and aggregated over 30 second incrementsfor more stable and meaningful time scales. By adopting thestandard loop detector aggregation scheme, emission statisticscan be directly correlated and used in the same way as thetraditional highway measures of flow, occupancy, and speed.In fact, they could be combined not only with VECTORhighway measurements but also any loop data, such as those

warehoused by PeMS.The final output component in the CalSentry system is a

remote user interface. A public website, utilizing the GoogleMaps API, was constructed to provide interested parties ac-cess to the emission measurements. Fig.2c shows the mapwith a color-coded view of a highway link. Using the samecolor scheme as the bins above, {yellow, orange, red}={low,medium, high}, the highway is colored to indicate the 30second aggregate emission value for the link. In this snapshot,the northbound direction is colored yellow indicating a lowemission level while the southbound is orange indicating amedium emission level. The map is similar to the morefamiliar navigation maps that have been color-coded forspeed. Similar emission coverage could be provided with moreCalSentry nodes to give a better sense of the current emissionconditions in a city or region.

Although not visualized, an important part of the CalSentrysystem is the database for historical record keeping. Withincreased coverage and data, the database will be valuable fordisplaying trends (e.g. the evolution of emission “hotspots” ina city over the course of a day) and as input to support largerscale emission modeling. This data will help transportationengineers and policy makers understand how commutes affectair quality and determine how to best manage or build futureroads.

The following sections describe the two main computationalcomponents in the CalSentry system. Section IV highlightsthe VECTOR modules and the emission/energy modeling ispresented in Section V.

IV. VECTOR TRAFFIC MONITORING

The following section describes highway traffic monitoringwith VECTOR [10]. The camera-based system is able todetect, track, and identify the type vehicles on the roadway.In addition, it produces traffic measurements, similar to tradi-tional loop detectors, in real-time.


(a) Sedan (b) Pickup (c) SUV (d) Van (e) Semi (f) Truck (g) Bike (h) Merged

Fig. 3: Sample images from each vehicle class.

FeatureExtraction

FCM DB

wkNN TrackClassification

Classification

ObjectDetection

Tracking

LDA

Csl

wc LT

Fig. 4: Block diagram for the VECTOR classification scheme.

A. Vehicle Detection and Tracking

Unlike a loop detector which is a spot sensor, camerasobserve a vehicle over a period of time while it travels throughthe camera’s FOV. Each video frame, every 33 msec, providesa new view which is used to describe the appearance of avehicle as well as its dynamics.

The VECTOR systems utilizes a single camera to monitorboth directions of a busy, 4-lane, highway. Vehicles are de-tected as moving regions using background subtraction andtracked using a global nearest neighbor optimization whichaccounts for both the dynamics as well as appearance. Thetrajectory of vehicle i,

Fi = {f1, . . . ft, . . . , fT }, with ft = [x, y, u, v]T , (1)

is the sequence of positions and velocities that describes thevehicle dynamics. The morphological appearance vector

Mi = [η0, . . . , η15]T = (2)

{area, breadth, compactness, elongation, perimeter, convexhull perimeter, length, long and short axis of fitted ellipse,roughness, centroid, the 4 first and second image moments}encodes the shape appearance of the particular vehicle.

B. Vehicle Classification

Using the appearance vector Mi, VECTOR classifies eachvehicle into one of the 8 different types, {Sedan, Pickup,SUV, Van, Semi, Truck, Bike, Merged}, seen in Fig. 3. Theclassification scheme is depicted in the block diagram of Fig.4. At a particular instant, Mi is transformed using lineardiscriminant analysis (LDA) and compared with a vehicledatabase using a weighted K nearest neighbor (wkNN) tech-nique) to produce class weights. The final vehicle label Li ofa track is determined after iterative refinement with each newvideo frame.

1) Feature Transformation: Appearance features were pro-jected using LDA [29] in order to separate vehicle classes andprovide a lower dimensional space to reduce computationalcomplexity for real-time implementation.

Let Dc = {x1, . . . , xNc} be a set of Nc training vectors for

class c, each of dimension d, with mean µc = 1Nc

∑Nc

i=1 xi.The full training set, D = {D1, . . . , DC}, is composed ofthe training samples from all classes and has mean µ =1N

∑Ni=1 xi, where N =

∑cNc. The LDA projection is given

by the maximization problem

PLDA = argmaxw

|wTSBw||wTSWw|

(3)

where SB is the between class scatter matrix and SW is thewithin class scatter matrix.

SB =

C∑i=1

Ni(µi − µ)(µi − µ)T (4)

SW =

C∑i=1

∑xk∈DC

(xk − µi)(xk − µi)T (5)

The solution to this maximization leads to the generalizedeigen problem SBw = λSWw. The top M = 5 eigenvectorsare retained to obtain the LDA projection matrix,

xLDA = PLDAx = [w1, ..., wi, ..., wM ]x (6)

2) Detection Classification: The wkNN rule is a mod-ification of the NN classifier for robustness to noise andoutliers that uses a soft assignment rule than a binary classmembership. The soft membership is denoted by the classweight

wc =

K∑i=1

xi∈Dc

1

‖ xi − xt ‖(7)

where a larger weight indicates a higher likelihood of a samplebelonging to class c based on the similarity to the K = 5closest training examples. In (7), xt is a new test sample toclassify given the training set consisting of all samples xi. Tocompletely specify the weights of the C = 8 vehicle types,K × C = 40 comparisons must be made.

3) Track-Based Classification Refinement: Information re-dundancy contained in sequential frames can be exploited toimprove vehicle type classification. A track-based refinementscheme is used to overcome measurement noise inherent ina single image. Given T images of a vehicle during track-ing, the track classification is found by maximum likelihood


05:00 07:00 09:00 11:00 13:00 15:00 17:00 19:000

200

400

600

800

1000

1200Friday 08/26/11 − Highway density

Time [24 hour]

Den

sity

[veh

icle

s/m

ile]

NorthSouth

(a) Density

05:00 07:00 09:00 11:00 13:00 15:00 17:00 19:000

20

40

60

80

100

120Friday 08/26/11 − Highway flow

Time [24 hour]

Flo

w [v

ehic

les/

30 s

ec]

NorthSouth

(b) Flow

05:00 07:00 09:00 11:00 13:00 15:00 17:00 19:000

10

20

30

40

50

60

70Friday 08/26/11 − Highway speed

Time [24 hour]

Spe

ed [m

ph]

NorthSouth

(c) Speed

Fig. 5: Density, flow, and speed for north and south bound directions of Interstate 5.

05:00 07:00 09:00 11:00 13:00 15:00 17:00 19:000

100

200

300

400

500

600

700Friday 08/26/11 − Highway south density of lanes

Time [24 hour]

Den

sity

[veh

icle

s/m

ile]

Lane 1 − Fast laneLane 4 − Slow lane

(a) Lane Density

05:00 07:00 09:00 11:00 13:00 15:00 17:00 19:000

10

20

30

40

50

60Friday 08/26/11 − Highway south flow of lanes

Time [24 hour]

Flo

w [v

ehic

les/

30 s

ec]


(b) Lane Flow

05:00 07:00 09:00 11:00 13:00 15:00 17:00 19:000

10

20

30

40

50

60

70

80Friday 08/26/11 − Highway south speed of lanes

Time [24 hour]

Spe

ed [m

ph]


(c) Lane Speed

Fig. 6: Individual lane density, flow, and speed for south bound direction of Interstate 5.

estimation.

L = argmaxc

T∑t=1

ln p(xt|c)

= argmaxc

T∑t=1

lnwt

c∑c w

tc

. (8)

The likelihood p(xt|c) of class c is approximated by normal-izing the instantaneous class distribution at time t defined by(7) in order to be valid probability. The track class label Lis refined each frame as a track is updated to leverage anyadditional appearance evidence before the trajectory ends.

C. Traffic Statistics

Using trajectory information, the time series of fundamentalhighway usage parameters, analogous to those obtained fromconventional loop detectors, is collected in real-time. TheVECTOR system delivers density (#vehicles

distance ), flow (#vehiclestime ),

and average speed (MPH) in 30 second intervals, averagedover a 5 minute window as shown in in Fig. 5.

1) Directional Measurements: The highway density (Fig.5a) indicates how crowded a roadway is and is computedby counting the number of vehicles in the camera viewnormalized by the roadway length. A loop detector cannotdirectly measure density because it is a spot sensor, insteaddensity is inferred based on occupancy and traffic flow. Traffic

flow (Fig. 5b) is a count of the number of passing vehicles in a30 second time interval. VECTOR produces the flow statisticby counting vehicle tracks as they exit the camera field ofview in a manner similar to loop detectors. The highway speedmeasure (Fig. 5c) is the average velocity of all vehicles seenin the 30 second interval. The roadway is calibrated basedon ground plane homography to convert pixels/sec imagetracking into MPH. VECTOR provides direct measurementwhile loop detectors often rely on algorithms based on flowand occupancy to estimate speeds [1].

2) Lane-Level Measurements: In Fig. 5 the traffic in northand southbound directions are compared during daylight hours.In this section of road, the southbound traffic is affected duringthe evening commute hours of 14:00-16:00. The causes canbe further investigated by focusing on the lane level trafficmeasurements presented in Fig. 6. During tracking, the lanenumber is determined based on position in the image. Thedensity of vehicles in the fast lane (lane 1) dramaticallyincreases during evening and results in a significant drop indriving speed.

3) Vehicle-Level Measurements: Since VECTOR is basedon video technology, rich contextual information not obtainedwith loop detectors can be extracted to further study trafficconditions. The loop-like traffic statics are further categorizedbased on the vehicle type as shown Fig. 7. Fig. 7a shows theproportion of vehicles on the road over the course of a day.


05:00 07:00 09:00 11:00 13:00 15:00 17:00 19:000

5

10

15

20

25

30

35

40Friday 08/26/11 − Highway south flow of vehicle types

Time [24 hour]

Flo

w [v

ehic

les/

30 s

ec]

sedanpickupsuvsemi

(a)

05:00 07:00 09:00 11:00 13:00 15:00 17:00 19:000

10

20

30

40

50

60

70

80Friday 08/26/11 − Highway south speed of vehicle types

Time [24 hour]

Spe

ed [m

ph]

sedanpickupsuvsemi

(b)

Fig. 7: Traffic statistics separated by vehicle type. (a) Distri-bution of highway flow by vehicle type. (b) Highway speedsof different vehicle classes

The speeds, shown in Fig. 7b, of sedans are greatest duringcongestion while SUV and pickup drivers are required toslow more with large commercial semi-trucks always travelingslowest.

V. EMISSION/ENERGY ESTIMATION

Current emphasis on environmental issues such as air pol-lution, greenhouse gases and energy consumption has fueledresearch in areas such as “green” vehicle technologies, alterna-tive fuels and ITS. This has resulted in the commercial successof hybrid automobiles and the re-introduction of consumerelectric vehicles as a way to curb emissions and energyconsumption. In the area of ITS, advances in traffic monitoringand data collection help improve traffic characterization andmanagement, however, it is still unclear exactly how the trans-portation network really affects emissions and how emissionscontributions can be characterized by location, time or mobilesource.

It is generally difficult to measure pollution from specificmobile sources under real-world conditions due to the mixingand dispersion of emissions as well as contributions fromadditional sources such as other vehicles, nearby factories andsecondary pollutants. These issues are further influenced byenvironmental factors such as meteorological conditions and

topographic features. Some methods of emission measure-ments, such as tunnel studies, remote sensing, and portableemission monitoring systems, address these problems to someextent. These methods, however, either do not isolate emis-sions from specific vehicles, are very limited in testing lo-cations or can not be practically applied to large sets ofvehicles. A method to better estimate emissions attributed totransportation and even specific vehicle classes in real-time isusefull for traffic management and policy makers as well ashealth organizations.

Using a vision based traffic management system, which isable to accurately track individual vehicles (collecting dynamicdriving patterns) and determine their type, it is possible toestimate the emissions and energy consumption from specificvehicles on the road. By accumulating emissions data overtime, a real-time map can be formed to indicate the level ofpollutants on our roadways.

A. Vehicle Class Emission Modeling

In order to accurately determine the amount of emissions orfuel usage from a particular vehicle, it is necessary to knowcertain vehicle characteristics such as weight, fuel type, enginedisplacement, after-treatment technology, vehicle model yearand vehicle age as well as how the vehicle is being operated(the driving profile). Unfortunately, it is not possible to deter-mine many of these vehicle characteristics using conventionaltraffic cameras. The resolution of these setups along with thevast number of vehicles on the road with varying character-istics makes this level of data collection almost impossiblewithout the use of other identifying techniques such as RF-tagsor license plate recognition. As shown earlier, it is, however,possible to distinguish between different classes of vehiclesusing conventional traffic cameras. Each class of vehicles hasdifferent emission properties which are generally related tovehicle size and type.

At the current time t, an instantaneous emission valueEpol(t) for pollutant pol can be estimated for each vehiclebased on the vehicle class, L, and dynamic profile, F (t)

Epol(t) = hF (L,F (t)). (9)

The functional mapping, hF , specifies how emissions areobtained from the vehicle information and must be specifiedor modeled.

B. Vehicle Specific Power Approach

There are various approaches to estimating vehicle emis-sions depending on the scope of the analysis and the availabledata. Traditional emission modeling techniques utilize averagespeed based emission rates for estimation. One of the fun-damental drawbacks of this modeling approach is that speedalone is not a good predictor of emissions since speed undervarious levels of acceleration will results in a wide range ofemissions. Acceleration is an important factor in the estimationof vehicle load which is well correlated with fuel use andconsequently emissions. In order to take advantage of thisadditional level of detail, VSP [11] was used as the basis foremission rates. VSP is defined as the instantaneous power to


TABLE I: VSP Parameter Approximation for VECTOR Vehi-cle Types

Type M [kg] Af [m2] Cr Cd

Sedan 1360 2.0 0.0135 0.34Pickup 2340 3.3 0.0135 0.43SUV 3035 3.44 0.0135 0.41Van 2270 3.46 0.0135 0.38Bike 230 0.65 0.0250 0.9Truck 11360 6.6 0.0094 0.7Semi 27300 10.0 0.0094 0.85

move a vehicle per the mass of the vehicle. The calculationfor VSP in kW/metric tons is based on the following equation,simplified from the power demand terms for a moving vehicle,

V SP (t) = v(t)(1.1a(t) + g sin(θ) + gCr)

+ρaCdAfv

3(t)

2M.

(10)

v = speed [m/s]a = acceleration [m/s2]g = gravity = 9.8 [m/s2]θ = grade [radians]

Cr = coefficient of rolling frictionρa = density of air = 1.2 [kg/m3] at sea level and 20◦ CCd = coefficient of aerodynamic dragAf = frontal area [m3]M = mass [kg]

The vehicle dynamic information [v(t), a(t)] are encoded inthe trajectory F (t). The parameter values in Table I are theapproximate values used for the seven VECTOR vehicle types(merged is excluded) based on the NCHRP 25-11 [30] dataset and values found in literature.

Using the VSP approach, emissions are estimated by mod-ifying (9) as

Epol = h(L, V SP (t)) (11)

where class L represents the VECTOR vehicle type categories(8) and V SP (t) encodes the vehicle emission-dynamics rela-tionships. The mapping, h, to produce emission values [31],is modeled based on CMEM and MOVES as described in thenext section.

C. Vehicle Specific Emission Tables

After accounting for the vehicle type and driving dynamicswith the VSP, the mapping h from VSP to a particularpollutant emissions is found. The VSP mapping is generatedusing two different simulation models, CMEM [12] and theEPA’s MOVES [32]. Emission tables developed for this projectprovide instantaneous emission rates for VSP values between0 and 40 kW/tonne and can be conveniently applied both inreal-time and in post processing. For each vehicle and at eachtime step, a VSP value is calculated using the equation (10)with vehicle class specific constants of Table I. An alternative,more computationaly intensive approach to estimating real-time vehicle emission would be to use individual vehicletrajectories directly as input to CMEM.

1) Comprehensive Modal Emissions Model (CMEM): TheVSP based emission tables for this project were primarilygenerated from modeling results from CMEM [12] which wasdeveloped at CE-CERT, University of California at Riverside.CMEM is a modal emissions model intended primarily for usewith microscale transportation models that typically producesecond-by-second vehicle trajectories. CMEM is capable ofpredicting second-by-second fuel consumption and tailpipeemissions of carbon monoxide (CO), carbon dioxide (CO2),hydrocarbons (HC), and nitrogen oxides (NOx) based ondifferent modal operations from an in-use vehicle fleet. CMEMconsists of nearly 30 vehicle/technology categories coveringlight-duty vehicles and Class-8 heavy-duty diesel trucks. WithCMEM, it is possible to predict energy and emissions from in-dividual vehicles or from an entire fleet of vehicles, operatingunder a variety of conditions.

One of the most important features of CMEM (and otherrelated models) is that it uses a physical, power-demandapproach based on a parameterized analytical representationof fuel consumption and emissions production. In this type ofmodel, the fuel consumption and emissions process is brokendown into components that correspond to physical phenomenaassociated with vehicle operation and emissions production.Each component is modeled as an analytical representationconsisting of various parameters that are characteristic ofthe process. These parameters vary according to the vehicletype, engine, emission technology, and level of deterioration.One distinct advantage of this physical approach is that it ispossible to adjust many of these physical parameters to predictenergy consumption and emissions of future vehicle modelsand applications of new technology (e.g., aftertreatment de-vices).

VSP and emission values are calculated for each CMEMvehicle category using the US06, FTP and MEC [30] drivingschedules. Vehicle population data from CARB’s EMFAC[8] model for San Diego County and calendar year 2010 isused to approximate fleet distributions for CMEM categories.CMEM categories are further grouped into the VECTORvehicle classes for compositing. Fig. 8 shows compositingresults for the VECTOR pickup class. In this figure the lightgreen lines show VSP emission results for individual CMEMvehicle categories within the VECTOR pickup class and theblack line shows the weighted composited VSP based emissionvalues for the VECTOR pickup class. The emission valuesare binned in 1 kW/tonne bins and in the figure the first binrepresents VSP values between 0 and 1 kW/tonne.

In addition to the VECTOR sedan, pickup and semi classes,specific van and SUV categories were developed to modelemissions from these two vehicle types with the CMEMmodel. In order to determine van and SUV CMEM categories,individual van and SUV vehicles from the NCHRP 25-11database from the original CMEM project were identified (20SUV vehicles and 37 vans), calibrated and modeled usingCMEM. The VSP based emissions from these vehicles werecomposited to create emission factors for those two vehicletypes specifically.

2) EPA MOVES Model: The remaining two VECTOR cate-gories, truck and motorcycle, are not supported by the CMEM


Fig. 8: VSP based emission rate values for the VECTORpickup class generated from weighted CMEM pickup truckcategories. Light green indicates the CMEM pickups and theblack is the average response used to generate the emissiontables.

model and are instead modeled using the 2010 MOVESdatabase which is the EPA’s latest mobile source emissionmodel. The VECTOR truck category is a broad category andencompasses a range of visually similar vehicle types suchas buses, garbage trucks, and medium heavy trucks. For themost part, these vehicles are large diesel engine driven vehiclesand for this project this class was approximated as an urbanbus according to EPA’s approximation for 1996-2006 class48 vehicles from heavy-heavy duty (HHD) vehicles [32]. Themotorcycle class is taken directly from the motorcycle baseemission rates found in MOVES.

The MOVES modeling methodology is based on VSPbinned emission rates. It is applicable at the microscale leveland can be integrated upwards for mesoscale and macroscaleapplications. The core of the MOVES modeling suite is aMySQL database which is referenced by the MOVES softwareand GUI to run elaborate analysis at various temporal andspatial resolutions. At the fundamental level, the MOVESmodel, is a set of emission and energy use tables binned byVSP operating mode. VSP operating mode bins are VSP binssplit not only by VSP, but also by mode such as acceleration,deceleration, braking, and speed range. MOVES VSP oper-ating mode bins are divided into 3 distinct speed ranges inan effort to separate emission speed effects. For this analysis,MOVES VSP operating mode bins with matching VSP ranges

were combined across vehicle speeds to create approximateVSP emission tables. An alternative approach is to use theMOVES VSP operating bins with the speed ranges. Emissionrates were extracted from the MOVES database by query usingthe appropriate sourceBinID for the regulatory class and asample model year group. The appropriate polProcessIDsfor CO, HC, NOx and total energy were used as well asageGroupIDs for 0-3 and 4-5 years. VSP operating modesbins between 11 and 40 were used. Pollutant emission factorswere queried from the emissionratebyage table and totalenergy was queried from the emissionrate table. Totalenergy was converted to CO2 using an oxidation factor of 1and carbon content of 0.00196 g/kJ [32].

VI. EXPERIMENTAL EVALUATION

The combination of real-time tracking and emissions model-ing present in the CalSentry system gives rise to a completelynew type of performance measurement system. In order toassess the performance, first the vehicle classification schemeis evaluated followed by a comparison with PeMS loop-basedemission output.

A. Visual Vehicle Type Classification

The vehicle classifier was only evaluated during the daylighthours of a single day because the detection and tracking doesnot work at night due to poor lighting and headlight reflectionon the road surface. Each hour, with sufficient sunlight, a 5minute video clip was saved for manual annotation of eachobserved vehicle. Both the type of vehicle and the lane oftravel were recorded, resulting in 6491 total tracks.

78.44% of the tracks were classified into the correct vehicletype. The full-day confusion matrix is presented in Table II.The classifier has difficulties with the Van and Truck classesbecause they are quite similar in appearance to SUV and Semirespectively. Table III gives the classification accuracy for eachhour of the experiment. The results from a single 5-minute clipis generally in the 80% range, except between 08:00-11:00.There is a significant performance drop during these morninghours due to adverse lighting conditions which caused largecast shadows from the vehicles. This typically caused largerdetections which resulted in misclassification into the SUVtype. Lighting issues are not new to visual monitoring andshadow suppression techniques [17], [33] could help improveclassifier performance.

B. Vision-based Traffic Statistics

Although real-time data is needed to understand current con-ditions, historical measurements provide the data for modelingand provides a deeper understanding of higher order effects.The observed flow and speed over a given week are shown inFig. 9. The differences between weekday and weekend trafficpatterns are quite clear. During the weekdays, there is a largeincrease in demand between the evening commute hours of15:00 and 17:00. The increased flow rate causes congestionand results in a large drop in speed. On the weekdays, thereis a 50% decrease in average speed during the commute hours


TABLE II: Confusion matrix for all hours of test. Total classification accuracy of 78.4% over 6491 test tracks

sedan pickup suv van semi truck bike mergedsedan 2726 127 202 55 0 0 1 0pickup 40 374 52 24 0 14 0 4

suv 411 113 1147 172 0 3 0 4van 15 11 54 83 0 6 0 7semi 0 0 0 0 26 1 0 1truck 1 5 1 2 11 36 0 0bike 1 0 0 0 0 0 18 0

merged 7 7 6 10 3 31 2 677total 3201 637 1462 346 40 91 21 693

% correct 85.2 58.7 78.5 24.0 65.0 39.6 85.7 97.7

TABLE III: Percentage Accuracy for Hourly Test Clips

time sedan pickup suv van semi truck bike merged total count06:21 94.9 59.5 81.9 31.6 50.0 33.3 0 97.5 81.5 40507:19 96.2 32.5 83.2 06.7 66.7 25.0 100 98.0 84.7 49708:17 61.2 33.3 91.9 14.3 50.0 50.0 100 96.7 67.6 53009:15 53.2 38.7 82.3 53.9 37.5 23.1 100 96.8 63.7 44410:13 36.8 26.7 77.3 26.7 71.4 40.0 0 93.9 51.0 35711:11 63.4 47.2 90.4 28.0 66.7 33.3 - 89.6 68.6 41712:09 86.0 71.7 82.6 48.0 50.0 37.5 100 96.9 80.8 43213:08 95.6 76.3 83.5 39.1 100 50.0 - 97.9 87.0 39314:06 96.9 77.8 84.2 18.2 - 66.7 100 94.9 86.2 44915:04 96.0 76.4 81.9 23.1 100 09.1 100 100 85.4 49216:02 97.1 66.2 76.0 24.0 100 55.6 100 100 85.7 55317:00 99.1 65.5 62.0 03.6 - 0 100 94.5 83.0 63017:45 89.0 75.9 52.8 10.0 - 100 67.0 97.7 76.0 29718:45 96.0 57.9 73.9 10.5 - 100 50.0 97.9 84.6 38219:43 95.0 77.8 78.5 0 100 100 - 100 86.5 222

while the weekends show no significant speed difference. Bystoring measurements in a database, they can be utilized tolearn and model variations in traffic patterns and behavior.

During the daylight test, 98% of vehicles traveling south(closest to the camera) were identified in the correct lane. Thenorthbound direction, which is much further away and suffersmore perspective distortion, only had 93% accuracy in laneassignment with the 3-lane performing worst at 84.4%. Thisimplies lane level accuracy is only possible with the propercamera-road configuration. However, link direction measure-ments are reliable even at sub-optimal camera configurations.

C. Real-Time Vehicle Emission Aggregation

Using Table I along with VSP-based emission profiles(highlighted in Fig. 8), the emissions from each vehicle werecalculated at 10 Hz and archived. The 10 Hz update rate waschosen as a compromise between VECTOR’s high video framerate of 30 fps and the natural 1 Hz operation of the microsim-ulation model. This compromise was necessary because at 30Hz the vehicle dynamics are noisier but at 1 Hz much ofthe trajectory profile would be lost. Emission measurementswere aggregated into 30 second increments before archival toprovide more stable and meaningful timescales which matchPeMS loop detector rates.

Fig. 10a presents the emissions (CO2, CO, HC and NOx)rates [g/sec] in the southbound direction of the highway. TheHC and NOx rates are scaled 50x and CO by 2x for plottingpurposes since the CO2 rate is much higher. The time-seriesplots have a number of spikes which should not come as asurprise because of the nature of traffic. During a particular30-second time interval, the number of and types of cars anddriving style, which greatly impacts emission production, isvariable. In Fig. 10b the emissions are aggregated over a

longer 5 minute time period. The emission measurements aresignificantly more stable at this time scale and provides a betterindication of the daily patterns. Around the 16:00-17:00 timeperiod there is a drop in emissions due to congestion. At thistime vehicles move slower and the VSP is greatly influencedby speed (0 speed results in no VSP output). Further workwill need to take into account idling emissions.

The emissions measurements generated by CalSentry arecompared with loop detector based estimates, shown in Figs.10c, 10d. These plots were generated by using the PeMSspeed measurement (no acceleration) in (10). The PeMS flowvalue was divided into the VECTOR vehicle classes based onregistration distribution resulting in fixed ratios of all vehiclesat all times. The 5 minute aggregates are of similar scale andhave the drop off during the evening commute, but the averagespeed version has a lower floor. As with CalSentry, the 30second aggregate is quite a bit more noisy (Fig. 10c) than the5 minute version.

CalSentry results are generally more variable and resultin slightly higher estimated emission rates than the loopsestimates based on average speed.

D. Future Work

Future work is needed to evaluate the validity of the emis-sion estimates, which will require more sophisticated tech-niques such as tunnel measurements or instrumented vehicles.Although it is expected that the accuracy of emission estimateswith the CalSentry system will be better than loop-basedestimates and that the CalSentry system will significantlyimprove the quality of on-road vehicle emission estimates,there are several factors which impact prediction accuracy.These factors include misclassification of vehicle categories,lumping of certain vehicle types (e.g. truck category), un-


05:00 07:00 09:00 11:00 13:00 15:00 17:00 19:000

20

40

60

80

100

120Highway south flow for different days in a week

Time [24 hour]

Flo

w [v

ehic

les/

30 s

ec]

MonTueWedThuFriSatSun

(a) Flow

05:00 07:00 09:00 11:00 13:00 15:00 17:00 19:000

10

20

30

40

50

60Highway south speed for different days in a week

Time [24 hour]

Spe

ed [m

ph]

MonTueWedThuFriSatSun

(b) Speed

Fig. 9: Traffic statistics for one week

05 07 09 11 13 15 17 190

5

10

15

20

25

30

35Friday 08/26/11 − Highway south emisison

Time [24 hour]

Em

issi

on r

ate

[g/s

]

CO

2 (1.0x)

CO (2.0x)NO

x (50.0x)

HC (50.0x)

(a) 30 sec CalSentry

05 07 09 11 13 15 17 190

5

10

15

20

25

30

35Friday 08/26/11 − Highway south emisison

Time [24 hour]

Em

issi

on r

ate

[g/s

]

CO

2 (1x)

CO (2x)NO

x (50x)

HC (50x)

(b) 5 min CalSentry

05 07 09 11 13 15 17 190

5

10

15

20

25

30

35

Time [24 hour]

Em

issi

on r

ate

(g/s

)

CO

2 (1.0x)

CO (2.0x)NO

x (50.0x)

HC (50.0x)

(c) 30 sec PeMS

05 07 09 11 13 15 17 190

5

10

15

20

25

30

35

Time [24 hour]

Em

issi

on r

ate

(g/s

)

CO

2 (1.0x)

CO (2.0x)NO

x (50.0x)

HC (50.0x)

(d) 5 min PeMS

Fig. 10: Highway emission estimates for 30-sec and 5-min aggregates.

known vehicle operating weights (especially for heavy dutyvehicles), unknown vehicle conditions (bad catalysts, tamperedemission controls, engine malfunctions, age of the vehicle),and the representativeness of emission factors which will varybased on the test vehicle sample size used to develop them.However, these same factors impact loop-based estimation andloops typically do not benefit from driving profiles or diversevehicle classification.

VII. CONCLUDING REMARKS

This paper introduced the first vision-based system fortraffic monitoring and emission/energy measurement calledCalSentry. CalSentry integrates live video processing, emis-sions modeling, and historical data archival into a visualizationframework for providing real-time emissions. An innovativesystem for real-time estimation of traffic emissions was de-veloped using a VSP-based approach which accounted forthe class of vehicle as well as the dynamic driving profileof velocity and acceleration. Using the CMEM and MOVESemission models, VSP-based emission profiles were generatedto convert the real-time vision tracking and VSP into CO2, CO,HC, and NOx emission estimates. A public website is providesa map that is color-coded based on the current emissionconditions on a highway link which could be extended forwider area coverage and used for policy decisions and real-time traffic management.

REFERENCES

[1] (2010) PeMS performance measurement system 10.3. Caltrans, TheUniversity of California, Berkeley, and PATH. [Online]. Available:http://pems.eecs.berkeley.edu

[2] B. T. Morris and M. M. Trivedi, “Trajectory learning for activity under-standing: Unsupervised, multilevel, and long-term adaptive approach,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 11, pp. 2287–2301,Nov. 2011.

[3] B. Morris and M. Trivedi, “Learning trajectory patterns by cluster-ing:experimental studies and comparative evaluation,” in Proc. IEEEConf. Computer Vision and Pattern Recognition, Miami, Florida, jun2009, pp. 312–319.

[4] B. T. Morris and M. M. Trivedi, “A survey of vision-based trajectorylearning and analysis for surveillance,” IEEE Trans. Circuits Syst. VideoTechnol., vol. 18, no. 8, pp. 1114–1127, Aug. 2008, Special Issue onVideo Surveillance.

[5] M. M. Trivedi, T. L. Gandhi, and K. S. Huang, “Distributed interactivevideo arrays for event capture and enhanced situational awareness,”IEEE Intell. Syst., vol. 20, no. 5, pp. 58–66, Sep. 2005.

[6] T. Gandhi, R. Chang, and M. M. Trivedi, “Video and seismic sensor-based structural health monitoring: Framework, algorithms, and imple-mentation,” IEEE Trans. Intell. Transp. Syst., vol. 8, no. 2, pp. 169–180,Jun. 2007.

[7] EPA, “Motor vehicle emission simulator (moves) 2010: User guide,”Office of Transportation and Air Quality: United States EnvironmentalProtection Agency, Tech. Rep., Dec. 2009.

[8] CARB, “Emfac2011 technical documentation,” California Environmen-tal Protection Agency Air Resources Board, Tech. Rep., Sep. 2011.

[9] R. Smit, L. Ntziachristos, and P. Boulter, “Validation of road vehicleand traffic emission models a review and meta-analysis,” AtmoshericEnvironment, no. 25, pp. 2943–2953, 2010.

[10] B. T. Morris and M. M. Trivedi, “Learning, modeling, and classificationof vehicle track patterns from live video,” IEEE Trans. Intell. Transp.Syst., vol. 9, no. 3, pp. 425–437, Sep. 2008.

[11] J. L. Jimenez-Palacios, “Understanding and quantifying motor vehicleemissions with vehicle specific power and TILDAS remote sensing,”Ph.D. dissertation, Massachusetts Institute of Technology, 1999.

[12] M. Barth, F. An, T. Younglove, G. Scora, C. Levine, M. Ross, andT. Wenzel, “Development of a comprehensive modal emissions model,”National Cooperative Highway Research Program, Tech. Rep., 1999.

[13] A. H. S. Lai and N. H. C. Yung, “Vehicle-type identification throughautomated virtual loop assignment and block-based direction-biasedmotion estimation,” IEEE Trans. Intell. Transp. Syst., vol. 1, no. 2, pp.86–97, Jun. 2000.


[14] V. Kastrinaki, M. Zervakis, and K. Kalaitzakis, “A survey of video pro-cessing techniques for traffic applications,” Image and Vision Computing,vol. 21, no. 4, pp. 359–381, Apr. 2003.

[15] D. Beymer, P. McLauchlan, B. Coifman, and J. Malik, “A real-timecomputer vision system for measuring traffic parameters,” in Proc. IEEEConf. Computer Vision and Pattern Recognition, San Juan, Puerto Rico,Jun. 1997, pp. 495–501.

[16] N. K. Kanhere and S. T. Birchfield, “Real-time incremental segmentationand tracking of vehicles at low camera angles using stable features,”IEEE Trans. Intell. Transp. Syst., no. 1, pp. 148–160, Mar. 2008.

[17] A. Prati, I. Mikic, M. M. Trivedi, and R. Cucchiara, “Detecting movingshadows: Algorithms and evaluation,” IEEE Trans. Pattern Anal. Mach.Intell., vol. 25, no. 7, pp. 918–923, Jul. 2003.

[18] J. C. Rubio, J. Serrat, A. M. Lopez, and D. Ponsa, “Multiple targettracking for intelligent headlights control,” in Proc. IEEE Conf. Intell.Transport. Syst., Sep. 2010, pp. 903–910.

[19] N. Buch, S. A. Velastin, and J. Orwell, “A review of computer visiontechniques for the analysis of urban traffic,” IEEE Trans. Intell. Transp.Syst., no. 3, pp. 920–939, Sep. 2011.

[20] S. Gupte, O. Masoud, R. F. K. Martin, and N. P. Papanikolopoulos,“Detection and classification of vehicles,” IEEE Trans. Intell. Transp.Syst., vol. 3, no. 1, pp. 37–47, Mar. 2002.

[21] J.-W. Hsieh, S.-H. Yu, Y.-S. Chen, and W. Hu, “Automatic trafficsurveillance system for vehicle tracking and classification,” IEEE Trans.Intell. Transp. Syst., vol. 7, no. 2, pp. 175–187, Jun. 2006.

[22] A. Ambardekar, M. Nicolescu, and G. Bebis, “Efficient vehicle trackingand classification for an automated traffic surveillance system,” in Proc.Signal and Image Processing, Aug. 2008.

[23] M. J. Leotta and J. L. Mundy, “Vehicle surveillance with a generic,adaptive, 3d vehicle model,” IEEE Trans. Pattern Anal. Mach. Intell.,no. 7, pp. 1457–1469, Jul. 2011.

[24] J. Wallace, D. Corr, P. Deluca, P. Kanaroglou, and B. McCarry, “Mobilemonitoring of air pollution in cities: the case of hamilton, ontario,canada,” J. Environ. Monit., no. 5, pp. 998–1003, 2009.

[25] J. Markey, “Ftp study background: a review of in-use driving surveys,”in FTP Review Project Public Workshop, Ann Arbor, MI, 1993.

[26] J.-Q. Li, W.-B. Zhang, and L. Zhang, “A web-based support systemfor estimating and visualizing the emissions of diesel transit buses,”Transportation Research Part D: Transport and Environment, no. 8, pp.533–540, 2009.

[27] K. Kraschl-Hirschmann, M. Zallinger, R. Luz, M. Fellendorf, andS. Hausberger, “A method for emission estimation for microscopictraffic flow simulation,” in IEEE Forum on Integrated and SustainableTransport. Syst., Jul. 2011, pp. 300–305.

[28] B. T. Morris and M. M. Trivedi, “Contextual activity visualizationfrom long-term video observations,” IEEE Intell. Syst., vol. 25, no. 3,pp. 50–62, 2010, Special Issue on Intelligent Monitoring of ComplexEnvironments.

[29] P. N. Belhumeur, J. P. Hespanha, and D. J. Kriegman, “Eigenfaces vs.Fisherfaces: Recognition using class specific linear projection,” IEEETrans. Pattern Anal. Mach. Intell., vol. 19, no. 7, pp. 711–720, Jul.1997.

[30] M. Barth, F. An, T. Younglove, C. Levine, G. Scora, M. Ross, andT. Wenzel, “The development of a comprehensive modal emissionsmodel,” National Cooperative Highway Research Program, Tech. Rep.,Nov. 1999.

[31] G. Scora, B. Morris, C. Tran, M. Barth, and M. Trivedi, “Real-timeroadway emissions estimation using visual traffic measurements,” inIEEE Forum on Integrated and Sustainable Transport. Syst., Jul. 2011,pp. 40–47.

[32] EPA, “Development of emission rates for heavy-duty vehicles in themotor vehicle emissions simulator (draft MOVES2009),” United StatesEnvironmental Protection Agency, Tech. Rep., 2009.

[33] A. Doshi and M. M. Trivedi, “Satellite imagery based adaptive back-ground models and shadow suppression,” Signal, Image and VideoProcessing, vol. 1, no. 2, pp. 119–132, Apr. 2007.

Brendan Tran Morris received his B.S. degreefrom the University of California, Berkeley in 2002and his Ph.D. degree from the University of Califor-nia, San Diego in 2010. His dissertation research on”Understanding Activity from Trajectory Patterns”was performed under Professor Mohan Trivedi andwas awarded the IEEE ITSS Best Dissertation Awardin 2010.

Morris’ research focus has been in real-time sens-ing and processing for understanding environmentsand situations. His interests include unsupervised

machine learning for recognizing and understanding activities, real-timemeasurement, monitoring, and analysis, and driver assistance and safetysystems.

Cuong Tran is a Ph.D. candidate at the ComputerVision and Robotics Research Laboratory, Univer-sity of California San Diego. His research interestsinclude vision-based human pose estimation andactivity analysis for interactive applications, intel-ligent driver assistance, human-machine interfaces,and behavior prediction. Tran received the B.S. incomputer science from Hanoi University of Tech-nology, Vietnam in 2004 and the M.S. in computerscience from UC San Diego in 2008. He is a VietnamEducation Foundation (VEF) Fellow.

Mohan Manubhai Trivedi is a Professor of elec-trical and computer engineering and the FoundingDirector of the Computer Vision and Robotics Re-search Laboratory and Laboratory for Intelligentand Safe Automobiles (LISA) at the University ofCalifornia, San Diego. He and his team are currentlypursuing research in machine and human perception,machine learning, human-centered multimodal inter-faces, intelligent transportation, driver assistance andactive safety systems. Trivedi serves as a consultantto industry and government agencies in the U.S.

and abroad, including the National Academies, major auto manufactures andresearch initiatives in Asia and Europe. Trivedi is a Fellow of the IEEE(“for contributions to Intelligent Transportation Systems field”), Fellow ofthe IAPR (“for contributions to vision systems for situational awareness andhuman-centered vehicle safety”), and Fellow of the SPIE (“for distinguishedcontributions to the field of optical engineering”).

George Scora earned a B.S. degree in Environ-mental Engineering in 1996 and an MS degree inChemical and Environmental Engineering (CEE) in2007 from the University of California, Riverside(UCR) where he is currently pursuing a PhD de-gree in CEE. He has over ten years of experienceworking at UCRs College of Engineering Centerfor Environmental Research and Technology as adevelopment engineer in the Transportation SystemsResearch group.

Mr. Scoras areas of interest include vehicle emis-sion modeling and transportation related air quality issues. He has been heavilyinvolved in the development of UCRs Comprehensive Modal Emissions Model(CMEM). As a student, Mr. Scora was also involved in work with the U.S.EPAs Office of Transportation and Air Quality in Ann Arbor, Michigan, whichfocused on analyzing emission data sets and helping to develop emissionfactors of both light-duty and heavy-duty vehicles for the MOtor VehicleEmission Simulator (MOVES) model.


Matthew J. Barth received his B.S. degree inElectrical Engineering/Computer Science from theUniversity of Colorado in 1984, and M.S. (1985)and Ph.D. (1990) degrees in Electrical and Com-puter Engineering from the University of California,Santa Barbara. Dr. Barth joined the University ofCalfornia-Riverside in 1991, conducting research inElectrical Engineering and the Center for Environ-mental Research and Technology (CE-CERT), wherehe is currently director.

Dr. Barth’s research focuses on applying engineer-ing system concepts and automation technology to Transportation Systems,and in particular how it relates to energy and air quality issues.

Dr. Barth is a member of the Institute of Electrical and Electronic Engineers(IEEE), Air and Waste Management Association (AWMA), TransportationResearch Board’s Transportation and Air Quality Committee, New Technol-ogy Committee, and ITS America’s Energy and Environment Committee. Hehas also served on several National Research Council (NRC) committees.Current research interests include Intelligent Transportation Systems, Trans-portation/Emissions Modeling, Vehicle Activity Analysis, Electric VehicleTechnology, and Advanced Sensing and Control.


05 07 09 11 13 15 17 190

5

10

15

20

25

30

Time [24 hour]

Em

issi

on r

ate

[g/s

]

CalSentryPEMSError

(a) CO2

05 07 09 11 13 15 17 190

1

2

3

4

5

6

Time [24 hour]

Em

issi

on r

ate

[g/s

]

CalSentryPEMSError

(b) CO

05 07 09 11 13 15 17 190

0.02

0.04

0.06

0.08

0.1

0.12

0.14

Time [24 hour]

Em

issi

on r

ate

[g/s

]

CalSentryPEMSError

(c) NOx

05 07 09 11 13 15 17 190

0.01

0.02

0.03

0.04

0.05

0.06

Time [24 hour]

Em

issi

on r

ate

[g/s

]

CalSentryPEMSError

(d) HC

Fig. 11: 5-Minute highway emission comparison between CalSentry and PeMS.

IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION …cvrr.ucsd.edu/publications/2012/Morris_ITS2012.pdf · vehicle proﬁles, the differences between different vehicle types and instantaneous

Documents