LOCATION-



Automated License Plate Readers Applied to Real-Time Arterial Performance:

A Feasibility Study

[pic]

Kevin Mizuta

Department of Civil & Environmental Engineering

University of Washington

June 7, 2007

TABLE OF CONTENTS

ABSTRACT 4

BACKGROUND 4

LOCATION & CONFIGURATION DETAILS 5

LICENSE PLATE READERS 9

PLATE READ PERFORMANCE 11

PLATE MATCHES 15

TRAVEL TIMES 19

REAL-TIME APPLICATION 25

CONCLUSION & SUGGESTIONS 28

REFERENCES 30

ACKNOWLEDGEMENTS 31

APPENDIX A: PIPS Technology Hardware Specification Sheets

APPENDIX B: WSDOT Base Plan Drawings

TABLE OF FIGURES

Figure 1: Map of Research Area (background from Mapquest) 5

Figure 2: Photo of reader installation at 80th Ave NE 6

Figure 3: Photo of West Half of Reasearch Area 8

Figure 4: Photo of East Half of Research Area 8

Figure 5: Photo of PIPS Technology P372 9

Figure 6: Photo of Reader Installation at 61st Ave NE 10

Figure 7: Graph of Eastbound 61st Ave NE Hourly Reads and Volumes 13

Figure 8: Graph of Westbound 61st Ave NE Hourly Reads and Volumes 13

Figure 9: Graph of Hourly Eastbound Volumes 14

Figure 10: Graph of Hourly Westbound Volumes 14

Figure 11: Map of Travel Time Segments (background from Mapquest) 16

Figure 12: Westbound Hourly Matches as Percentage of Reads 17

Figure 13: Eastbound Hourly Matches as Percentage Reads 17

Figure 14: Map of Match-to-Read Percentages (background from Mapquest) 18

Figure 15: Graph of Travel Times Sample (Raw) 20

Figure 16: Graph of Travel Times Sample Smoothed Once 20

Figure 17: Graph of Travel Times Sample Smoothed Twice 21

Figure 18: Graph of Segment 1 Travel Times 21

Figure 19: Graph of Segment 2 Travel Times 22

Figure 20: Graph of Segment 3 Travel Times 22

Figure 21: Graph of Segment 4 Travel Times 23

Figure 22: Graph of Segment 5 Travel Times 23

Figure 23: Graph of Segment 6 Travel Times 24

Figure 24: Graph of Matches per Interval by Hour 26

ABSTRACT

This report will analyze the license plate data collected and archived by an arterial monitoring system made up of automated license plate readers. The system will be installed by the Washington State Department of Transportation on SR-522. The data will be analyzed for accuracy and reliability. Based on the performance of the system the feasibility of using the set-up for a real-time traffic information system will be examined.

BACKGROUND

Travel times have become a popular performance measure for traffic conditions in the transportation industry. With increasing pressure for public agencies like the Washington State Department of Transportation (WSDOT) to provide reports of accountability, performance measures have become more important than ever for both policy makers and the public. Travel times include origin, destination, and everything in between, which encompasses all factors that can influence travel- congestion, incidents, weather, construction, etc. Travel times are also popular because time is a measure that everyone in the general public can grasp and understand.

In the greater Seattle area, freeway travel times have become a popular measure during the morning commute. The major television news networks, KIRO, KING, KOMO, and KCPQ all report travel times during their morning news shows provided by WSDOT. WSDOT has also successfully launched a Travel Time VMS system that automatically displays travel times on variable message signs in key freeway locations when they are sitting idle. WSDOT is interested in expanding their travel time monitoring to arterial corridors.

The method used to estimate travel times on freeway routes involves occupancy data from loop detectors generally spaced every half mile. The occupancy of one point along a segment is assumed to represent the condition for that entire segment. This data is converted to speed which then is used to calculate the amount of time needed to travel that segment. Continuous segments are added up along a given route to produce a travel time. Using this same method to produce travel times along an arterial corridor is a problem. Arterials by definition include signalized intersections that produce delay and therefore produce occupancy at regular intervals. The assumed constant speed along segments to utilize occupancy data does not apply to arterials which has interrupted flow. The recent development and availability of automated license plate reader technology to the transportation industry may help solve this problem.

This system has the potential to be useful for several different audiences or interests. The data can be presented as real-time traffic information, but also can be used for performance measures at the engineering and administrative level. This report will examine if and how current license plate reader technology could help for such a system.

LOCATION & CONFIGURATION DETAILS

The location of the research corridor is a 2.35 mile stretch of SR522 around the northern tip of Lake Washington in the greater Seattle area. The corridor starts at NE 170th St (mp 5.64) in the city of Lake Forest Park to 80th Ave NE (mp 7.99) in the city of Kenmore. This location was chosen because it coincides with WSDOT’s Arterial Traffic Management System (ATMS) on the same roadway. The ATMS allows a remote station at the Traffic Management Center to monitor and adjust timing of the corridor traffic signals which are all maintained and operated by WSDOT. By co-locating the readers along the same corridor, they share the State’s existing fiber optic network used for the ATMS communications. 6 readers were installed at 3 intersections (one for each direction of SR522) shown in Figure 1 below. The base map drawing for the corridor is available in the Appendix.

[pic]

Figure 1: Map of Research Area (background from Mapquest)

The readers at each end of the corridor (170th St NE and 80th Ave NE) were installed in October 2006. The readers at 61st Ave NE were installed at the end of April 2007 just before analysis for this report began. The readers were installed on existing mast arms of the intersection signal heads. The readers were positioned to read license plates from the rear of vehicles as they passed under. This “rear firing” position from above is the manufacturer’s suggested configuration for optimal performance. At 61st Ave NE, the long reach of the mast arm over the eastbound lanes allowed the cameras for both directions to be mounted on the same arm. The westbound reader at 80th Ave NE could not be installed like the others due to the limitations of existing conduit runs. The reader was installed on the mast arm above the eastbound lanes and aimed across the roadway to read plates from the front of vehicles as they approached on the westbound side. Figure 2 below is a photograph of the mast arm at 80th Ave NE over the eastbound lanes.

[pic]

Figure 2: Photo of reader installation at 80th Ave NE

SR522 (Bothell Way NE) is a 6-lane arterial through most of the corridor, where the 4 inside lanes are general purpose and the 2 outside lanes are transit only. The transit lanes end just west of 80th Ave NE where Bothell Way necks down to 4 lanes. All of the westbound readers were positioned to read vehicles traveling in the inside lane, closest to the median. All of the eastbound readers were positioned to read vehicles traveling in the outside general purpose lane.

The west end of SR522 meets I-5 in Seattle as Lake City Way then runs north to NE 145th St where it becomes Bothell Way and continues through Lake Forest Park,

Kenmore, and Bothell. SR522 gives motorists a non-freeway option to travel around the north end of Lake Washington and links I-5 on the west side to I-405 on the east side. SR522 continues eastward beyond I-405 to connect the Cities of Woodinville and Monroe then ends at SR2. Bothell Way sees well over 50,000 vehicles per day along this corridor with a morning peak in the westbound direction and an evening peak eastbound.

The 2 halves of the corridor split by the intersection at 61st Ave NE have very different characteristics. The west half is just shy of 1 mile long and runs close to the northern shore of Lake Washington (see the map of Figure 1). Bothell Way serves as a northern boundary for the prime residential property nestled between it and the lakeshore. There is limited access into these neighborhoods served primarily by a couple of controlled intersections. The only area of high activity for this segment is the Lake Forest Park Center located on the north side of SR522 between NE 170th St and Ballinger Way NE. This segment generally has a very small count of uncontrolled side streets and driveways. Figure 3 below shows a photo from the middle of this segment taken from the transit lane in the westbound direction.

The eastern half of the study corridor has a much different character. Here, Bothell Way NE serves as a boundary separating the lakeshore industrial zone on its south from the extremely active commercial/retail area on its north. By design, there is still limited access to the areas on the south to minimize conflicts with the Burke Gilman Trail (used heavily by pedestrians and bicyclists) that runs along the south edge of the street. There is a large number of uncontrolled accesses to the businesses on the north side. Figure 4 shows the density of businesses along the north side of the segment. At 68th Ave NE, SR522 continues eastward away from Lake Washington following the Sammamish River through Bothell.

The differences in the characteristics of these segments are important in understanding the data results from this research. The characteristics had a tremendous influence on the ability to produce travel times. Also, the westbound reads coming from the inside lane while the eastbound reads coming from the outside lane seemed to make an impact on the analysis.

The volume data for this report was provided by the ATMS system which archives volume and occupancy data from the signal loops every 5 minutes. The ATMS system manages 7 signals on Bothell Way, 6 within the study corridor:

SR522 @ NE 165th St (outside study area limits)

SR522 @ NE 170th St

SR522 @ Ballinger Way NE

SR522 @ 61st Ave NE

SR522 @ 68th Ave NE

SR522 @ 73rd Ave NE

SR522 @ 80th Ave NE

The mainline advance detection loops approaching the signals provided useful volume data for analysis. Details of which specific loops were used for counts are given in the analysis sections of this report.

The study period used for this analysis was Tuesday through Thursday within May 1st and May 10th of 2007 (six 24-hour periods). This was the only period when all 6 readers were installed and functioning at the same time. Mondays, Fridays, and weekends were omitted from averaging due to their significantly different traffic patterns. The license plate reads and volume data were pulled from these same 6 days.

[pic]

Figure 3: Photo of West Half of Research Area

[pic]

Figure 4: Photo of East Half of Research Area

LICENSE PLATE READERS

The automated license plate recognition (ALPR) cameras used for this report are made by PIPS Technology. PIPS Technology specializes in developing components and systems for plate recognition applied to law enforcement, tolling, parking, and intelligent transportation systems. PIPS Technology cameras are widely used in Europe and in 2006 started to gain momentum in the US market. The model used in this research is the P372 purchased from PIPS Technology in October 2006 (the current P372 is packaged in a new model called the Spike +). The following is the manufacturer’s system identification:

VES (Violation Enforcement System)

Version 3, release x06, build 168

Running on C EXECUTIVE 3.3

[pic]

Figure 5: Photo of PIPS Technology P372

The P372 is an IP addressable, self contained recognition system where the camera, illuminator, and processor are housed in the same unit. The illuminator is infra-red allowing the cameras to pick up license plates in low light and nighttime conditions. Figure 5 shows a close up view of the P372 camera mounted for westbound traffic at NE 170th St. Every P372 camera is mated to an external breakout box. This box supplies DC current to the camera from an AC power source and serves as a protective switch for the unit from external electrical surges and lightning. Manufacturer’s spec sheets for both the P372 and the breakout box are available in the Appendix. The hybrid communication/power cable running from the camera to the breakout box is supplied by the manufacturer at 15 feet long. For most of the installations, the box was located on the mast arm along with the camera or on the utility pole supporting the mast arm. Figure 6 shows the installation at 61st Ave NE where the boxes are located on the mast arm between the readers. At each intersection, the cabinet housing the signal controller is located on a concrete pedestal in the southeast quadrant (the fiber optic communication line runs along the south side of the street). Cables for power and communication run from the breakout box to the existing service in the signal cabinets sharing the signal system conduits. The manufacturer suggests these runs should be less than 250 feet. This is the primary reason the westbound camera at 80th Ave NE could not be positioned like the others. The existing conduit runs would require a significantly longer run from the breakout box to the signal cabinet.

[pic]

Figure 6: Photo of Reader Installation at 61st Ave NE

Each P372 has a small amount of internal memory. As the cameras made plate reads, it stored the data (plate number, date, time, etc.) in internal files. When the files reached a certain variable size they were pushed to a centrally located server at the Traffic Management Center. For this study, the variable size was set to the minimum of 10 kilobytes. This resulted in varying time between file transfers based on the current traffic intensity.

The P372 was first introduced to WSDOT in late 2006 as a solution to temporarily collect travel times for performance measures. The cameras are set up on temporary tripods at roadside to collect license plate reads at 2 separate points using a laptop at each camera to log the data. Before and after travel times are compared to measure the effectiveness of roadway improvements and/or document the delays actual construction activity causes. The successes of these temporary stand-alone applications have spawned interest in the feasibility of a more permanent networked system studied in this report.

PLATE READ PERFORMANCE

In order to analyze plate read performance, the plate reads were aggregated into hourly groups for each day and averaged for the study period (plate reads were archived in files of same size but different time intervals because they were pushed to the server based on file size, not time interval). The average hourly reads over the 6-day period were compared to the average hourly volume from the closest available ATMS loop count. Table 1 shows the reads as a percentage of volume for each camera. The percentages are based on the average daily numbers. The table also includes where the volume count was taken for each camera.

Table 1: Plate Reads as Percentage of Local Volume

|Eastbound |Volume Count Location |% of Volume |

|NE 170th St |NE 170th St (upstream) |93% |

|61st Ave NE |68th Ave NE (downstream) |77% |

|80th Ave NE |80th Ave NE (upstream) |79% |

| | | |

|Westbound |Volume Count Location |% of Volume |

|80th Ave NE |80th Ave NE (upstream) |68% |

|61st Ave NE |Ballinger Way NE (downstream) |98% |

|NE 170th St |NE 170th St (upstream) |97% |

One factor that would lower percentage is that plate reads and volume counts were not taken from the exact same location. Advance loop detectors for an intersection could be placed anywhere from 150-300 feet behind the stop bar and the cameras were reading vehicles anywhere from 50-100 feet after. This distance is even greater when the volumes are taken from the closest neighboring intersection such as the case for 61st Ave NE in either direction. This is enough distance for many vehicles to change lanes or leave the corridor, and new vehicles to enter.

The resulting read-to-volume percentages came as no surprise. The westbound reader at 80th Ave NE had the worst percentage of all 6 readers due to its compromised installation. The camera was not reading from above and behind as suggested by the manufacturer for optimal performance. The camera was aimed at the front of vehicles at an angle which would compromise field of view, contrast, and reflectivity. The camera did not successfully read as high a percentage of the actual vehicles traveling in the target lane as in other locations.

Westbound percentages (except 80th Ave NE as explained above) were higher than eastbound percentages because of the difference in target lanes. Figure 7 and 8 show the graph of the hourly reads and volumes for each direction of 61st Ave NE. The westbound readers were reading the inside lane where the potential for vehicles to move out of the lane or leave the corridor is less. Based on general experience, motorists tend to travel in the inside lane to continue moving through an intersection or to turn left. The eastbound readers were aimed at the outside lane where there is a much higher potential for vehicles to change lanes (to the right or left). When vehicles want to exit right, they tend to travel in the target lane to stay out of the transit lane until it is absolutely necessary. This also explains why the western reader percentages are higher for both directions because there is less opportunity for vehicles to exit the corridor and less potential for this section to be a trip destination. During the morning peak the westbound reads plot greater than the actual reads (Figure 8). The volumes counts for the reader at 61st Ave NE were taken downstream from the reader at Ballinger Way NE. There is a heavy movement of morning commuters turning at Ballinger Way to head towards I-5. Several vehicles are being read in the inside lane at 61st but moving to take a right at Ballinger before reaching the count loop.

It is safe to say that the performance of the license plating reading systems themselves is not an issue. The ALPR cameras performed fine with a high read-to-volume percentage when outside factors are minimized. All the read-to-volume percentages would be in the 90th percentile if they were configured similarly to the westbound readers at 61st Ave NE and NE 170th St- rear firing from above in the inside lane.

The volumes were taken separately for each location instead of using an overall corridor volume because the volumes fluctuated due to the movement in and out of the corridor from driveways and side streets. During the morning peak, the volumes grew larger as you traveled down the corridor in either direction as more commuters entered from the side streets to travel to work heading west to I-5 or east to I-405. The exact opposite was occurring during the evening when commuters were headed back to the area. This shows how much Bothell Way NE is utilized by the local residences as the main artery in and out of area. Figure 9 and 10 show the hourly volumes for eastbound and westbound respectively. All volume counts represent one lane of travel corresponding to the same lanes as the cameras.

[pic]

Figure 7: Graph of Eastbound 61st Ave NE Hourly Reads and Volumes

[pic]

Figure 8: Graph of Westbound 61st Ave NE Hourly Reads and Volumes

[pic]

Figure 9: Graph of Hourly Eastbound Volumes

[pic]

Figure 10: Graph of Hourly Westbound Volumes

PLATE MATCHES

Travel times are calculated by matching plates from 2 separate readers in the same direction and taking the difference of the time stamps. A small program was developed by WSDOT’s ITS Software group to match the plates of the archived ALPR data. This manually executed program asks for a separate upstream file and downstream file. These files contain the plate reads for the period of time the user is interested in matching plates. The program looks for matches and outputs a csv file with one match per row and columns for plate number, start date, start time, end date, end time, and travel time. In a real-time information system, this action would occur automatically as each plate read reached the server or in some predetermined time interval. More details on the architecture of a real-time system are discussed in the REAL-TIME APPLICATION section.

There were 2 common events that resulted in inaccurately long travel times. There were several matches where a vehicle would hit an upstream reader on day 1 but not hit the downstream reader until a separate trip on day 2. This resulted in false travel times well above 20 hours. Also vehicles would temporarily stop along Bothell Way at the gas station, coffee shop, or grocery store, etc then continue down the corridor. This resulted in travel times with varying degree of error. To reduce these events, all travel times above 30 minutes were automatically omitted from analysis. 30 minutes was more than enough time to encompass any possible delay used in a real-time information system for the 2.35-mile long stretch.

In order to analyze the different travel times, the corridor was split and numbered as 6 separate travel time segments. Table 2 identifies the segments and their corresponding upstream and downstream readers and Figure 11 shows how they fit geographically on a map.

Table 2: Travel Time Segments

| |Description |Upstream |Downstream |

|Segment 1 |entire westbound |80th Ave NE |NE 170th St |

|Segment 2 |westbound, east portion |80th Ave NE |61st Ave NE |

|Segment 3 |westbound, west portion |61st Ave NE |NE 170th St |

|Segment 4 |entire eastbound |NE 170th St |80th Ave NE |

|Segment 5 |eastbound, west portion |NE 170th St |61st Ave NE |

|Segment 6 |eastbound, east portion |61st Ave NE |80th Ave NE |

Segments 2 and 3 are sub-segments of Segment 1, and Segments 5 and 6 are sub-segments of Segment 4. The plate reads for each location for the 6-day period were processed through the matching program and aggregated into hourly counts. These counts were compared to the volume counts of the downstream detectors. In a real-time environment the server would receive a plate read from a detector then look at reads from the upstream detectors to find matches (see REAL-TIME APPLICATION). A good measure of matching performance would be the ratio of matches found compared with the number of plate reads collected at the downstream detector for each segment.

[pic]

Figure 11: Map of Travel Time Segments (background from Mapquest)

Figure 12 and 13 show graphs for the hourly matches as percentages of downstream plate reads for eastbound and westbound, respectively. The graphs also give the average match-to-read percentage for each segment. For either direction, the percentages were poor for the long spans (14% for Segment 1, 18% for Segment 4). This is justifiable because the longer spans provide more distance and opportunity for vehicles to move out of the target lane. Also, there are many vehicles entering Bothell Way from the side streets and driveways. This would give the downstream reader a larger number of plates and less potential for matches with the readers upstream. I believe Segment 1 would have percentages higher than Segment 4 if the westbound reader at 80th Ave NE was installed in the optimal configuration. The fact that this reader is only capturing 68% of the vehicles passing reduces the chances of successful matches with readers upstream.

Taking a closer look at the westbound direction, NE 170th St matched 14% of its reads with 80th Ave NE (Segment 1), but matched 44% of its reads with 61st Ave NE (Segment 3). The western segment did more than 3 times better than the whole corridor because of the limited access for this section and the upstream and downstream readers capturing averages of 98% and 97% of the volume, respectively. Looking at the eastbound direction, Segment 5 had a 40% match-to-read ratio because it shares the same characteristics as Segment 3 mentioned above. Although the reader at 61st Ave NE captured only 77% of the volume, the 93% capture rate of the upstream reader kept the match percentage high. Figure 14 shows how all the matches-to-reads percentages come together geographically.

Segment 2 westbound and Segment 6 eastbound had similar percentages (20% and 24% respectively) because they share the same eastern end of the corridor. The heavy retail activity in this section adds potential lane changes and entering/exiting vehicles. Again, I believe the westbound segment would have performed better than eastbound based on the target lane if the westbound reader at 80th Ave NE was installed in the optimal configuration.

[pic]

Figure 12: Westbound Hourly Matches as Percentage of Reads

[pic]

Figure 13: Eastbound Hourly Matches as Percentage Reads

[pic]

Figure 14: Map of Match-to-Read Percentages (background from Mapquest)

TRAVEL TIMES

In order to minimize the effects of the outlying travel times caused by drivers stopping for reasons other than traffic, the data was smoothed with an averaging process. If a travel time was more than one standard deviation above or below the moving average of the 10 previous entries, the travel time was removed. Two iterations of this process were executed for each segment.

Figures 15 through 17 all plot the travel times from the same time period and location (Segment 1, noon to 8pm) to show the smoothing affected the numbers. Figure 15 is a plot of the raw travel times for the sample period. The black line shows the moving average based on a period of 10 previous points. Figure 16 shows the same data after one iteration of the smoothing process. Naturally, most of the omitted data comes from the top of the graph because travel time variation tends toward values greater than actual, not lesser. Variation toward the lower end only occurs in the rare occasion that a vehicle significantly speeds through the entire corridor. Figure 17 shows the same data after 2 iterations. All the outliers were virtually eliminated. The iterations stopped at 2 because there was little change to the moving average with a third iteration.

Figures 18 through 23 show the smoothed travel times and moving average for each of the 6 segments. On the day scale, the variance in travel times coincides perfectly with the volume/congestion periods for the corridor. The system, at all segments, works to accurately capture the relative traffic condition based on the travel time measure. The smaller oscillations that occurred within each hour are caused by the variance in total stop time each vehicle experiences when driving through the network of signalized intersections (what can be considered the variation within control). The low point of these oscillations would be the travel time at that period for the lucky vehicle that spent the least amount of time stopped at lights while the high point would be the travel time of the unluckiest vehicle.

The shorter segments give better detail to the location of delay for the corridor. Segment 1 (Figure 18) shows two distinct peaks in travel times for the two daily rushes. By looking at Segment 2 and 3 (Figures 19 and 20) we can see that the delay for the morning peak is caused primarily by Segment 3 (the western half) while the delay for the evening peak is caused primarily by Segment 2 (the eastern half). This seems justified because the eastern half is much busier with commercial/retail activity. The evening work-to-home trips are mixing with the other home based trips that are constantly building throughout the day. Segment 2 is also upstream from Segment 3 so the delays in 2 help conditions in 3.

In the eastbound direction, most of the delay was happening in Segment 5 (Figure 22). Segment 6 (Figure 23) had somewhat constant conditions but improved significantly during the evening commute when Segment 5 upstream had the most delay. The upstream congestion was improving traffic conditions downstream.

[pic]

Figure 15: Graph of Travel Times Sample (Raw)

[pic]

Figure 16: Graph of Travel Times Sample Smoothed Once

[pic]

Figure 17: Graph of Travel Times Sample Smoothed Twice

[pic]

Figure 18: Graph of Segment 1 Travel Times

[pic]

Figure 19: Graph of Segment 2 Travel Times

[pic]

Figure 20: Graph of Segment 3 Travel Times

[pic]

Figure 21: Graph of Segment 4 Travel Times

[pic]

Figure 22: Graph of Segment 5 Travel Times

[pic]

Figure 23: Graph of Segment 6 Travel Times

REAL-TIME APPLICATION

The current ALPR setup is slightly different than what is needed for a real-time traffic information system. Most of the difference lies in the way the data is transferred from the cameras to the server. As mentioned in the LICENSE PLATE READERS section, the cameras push data to the server when the files containing the plates reach a certain size. Even at the minimum file size setting (10 kb) this equals about 240 plate reads. Time between file transfers can be hours at night compared to minutes during peaks. Although this file transfer method works for analyzing archived data for this report, it does not allow for a quick response to traffic conditions in real-time. Plate reads must get to the central server in small, constant time intervals. In order for the system to produce useful real-time travel information, 3 major steps must occur:

• First, the server needs to receive the plate reads from the cameras at a constant rate or immediately throughout the day. The server could poll the cameras at set time intervals, taking what is collected at the camera or the server could receive plates from the cameras as they are read through an open socket connection. The plate reads would then be archived at the server.

• Second, a program at the server would take archived downstream plates and attempt to find matches with the archived plates from the corresponding locations upstream. If the server was polling the cameras at set intervals, the program could run its process on the plates in each packet as they were received. If there was an open socket connection and plate reads were coming in individually, the program would run at set time intervals using plates collected during that interval. The size of these intervals would determine the resolution at which travel times would be outputted. The program would then throw out the outliers, and output a valid travel time for that time interval

• Third, the travel times would need to be displayed or represented in a format that would be useful to engineers and/or the public. This would most likely be done graphically on a web-based interface. Real numbers could be shown or colors representing predetermined conditions could be used similar to the WSDOT Seattle Area Traffic Flow Map () which shows colors based on occupancy data.

With WSDOT’s current network, the server would be set up to receive plates from the cameras as they are read. Each plate read would be stored in a separate file for each location at the server. The suggested time interval to execute the plate matching process for this corridor is 4 minutes. This interval was chosen based on the idea that at least 10 matches should be collected to make a good travel time average. This number could be debated but a minimum of 10 was chosen to stay consistent with the previous data processing in this report. 4 minutes was determined by looking at the matches collected between westbound NE 170th St and 61st Ave NE (Segment 3). The performance of this segment represents the potential for readers in this corridor set up in the ideal configuration. Figure 24 shows matches by 2, 3, 4, and 5 minute intervals based on hourly data for Segment 3. If we temporarily ignore the after-hours period, 4 minutes allows all intervals to be above a 10 match average through both peaks. The number of average matches drops below 10 at around 8pm. A larger interval could be chosen if there is a need for travel times in the later hours.

[pic]

Figure 24: Graph of Matches per Interval by Hour

There are not enough matches during the after-hours for the same process to generate reliable travel times. Although Segment 3 is averaging over 4300 matches per day, only 55 of those are between the hours of 12am to 5am (.733 matches per 4-minute interval). Evidently, this is also the period when real-time traffic information is least valuable. One way would be to require each interval to have the minimum match count value (in this case 10). If the interval does not meet the minimum, the output would be the preset ideal travel time (calculated using historic data). This could be set to only work during certain hours or throughout the day.

A ceiling travel time should be determined to throw out the travel times outside feasibility such as the matches caught between two separate days. This ceiling time was 30 minutes for the analysis in this report (all travel times above 30 minutes were omitted from analysis). 30 minutes is a good starting point for the 2.35 mile corridor to eliminate the “unreal” travel times but still capture all potentially valid times. This ceiling travel time can be designed into the matching program by limiting the matching process to only look at upstream plates within a set “look back” period.

Smoothing should be accomplished within each 4 minute period. All travel times outside of one standard deviation above and below the mean should be omitted. Sufficient numbers resulted when this process was completed for 2 iterations (see TRAVEL TIMES section).

The following steps show how the travel time data processing would occur during the morning peak at one location given the suggested parameters:

1. All 6 cameras have an open socket connection with the server and are sending plate reads as they are made.

2. At some 5 minute interval during the morning peak, the server takes the plates it has collected from the last 5 minutes for NE 170th St and attempts to find matches with the next upstream reader using plates archived within the last 30 minutes.

3. After the matches are found and travel times extracted, the match count is checked to meet the 10 count minimum. If the minimum is not met, the predetermined best time based on historic data is used. If the minimum count is met, the travel times are averaged and the outliers are removed (2 iterations). The average travel time resulting is the accepted travel time for the interval.

4. The travel time is then sent to a web server to be presented on a web-based GUI.

The results from this data analysis show that a real-time travel information system could be developed with the ALPR readers. The resolution at which travel times are reported would be directly linked to the volume of the travel time segment. A corridor like this one which sees over 50,000 vehicles daily can produce travel times every 3-5 minutes with one lane targeted. A much higher resolution (smaller intervals) could be achieved if every lane of an arterial was covered.

CONCLUSION & SUGGESTIONS

The analysis shows that ALPR technology has the potential to play a big part in arterial performance measures by producing travel times. The data show a useful system can be set up for an arterial even with one lane targeted in each direction. Plate readers installed in an ideal configuration could produce travel times at a resolution that would be useful for real-time traffic condition reporting for engineers and the public. The following is a list of takeaways collected from the analysis:

• The license plate readers should be installed in the optimal configuration with all outside factors that diminish performance minimized. Plate reads to volume percentages of 97% and 98% were achieved in this corridor when the ideal configuration was paired with the ideal target lane.

• For the PIPS Technology P372 camera, the rear firing from above position should be used (manufacturers suggested position). Although the camera is advertised to work in multiple configurations, there is significant performance loss when changed from optimal. Our “compromised” reader made only 68% of the plates in its lane (a level of performance that would not allow us to deploy the proposed real-time system). The resolution of travel time data is directly linked to the camera’s plate reading performance.

• A ceiling travel time value (the value of 30 minutes was used as a conservative threshold for this report) should be determined primarily to eliminate the reads that span more than one day, but also to eliminate some of the false travel times due to vehicles stopping for reasons other than traffic. The remaining false travel times will be eliminated in the smoothing process.

• The travel time segment with the ideal reader configuration targeting the ideal lane (Segment 3) had a match-to-read ratio of 44%- meaning 44% of the plate reads were matched for travel times. The same segment in the other direction (Segment 5) targeting a non-ideal lane had a match-to-read percentage of 40%. This may indicate a standard percentage in the 40th percentile for a 6-lane arterial with one targeted lane.

• For either direction, the plotted travel times for the smaller segments gave more information than the aggregated large segment. By cutting each large segment in half, it increased the number of successful matches by closing in the distances from upstream and downstream readers. This reduced the potential in each segment for vehicles to enter/exit and change lanes. The smaller segments also told which half the delays in the large corridor were occurring. This would be more useful for a large corridor made up of many smaller segments. The number of segments would depend on the purpose of the traffic data. Smaller segments would be useful for operations but large segments are sufficient for public travel information.

• The suggested resolution for the research corridor for travel times is every 4 minutes. This is based on the idea that at least 10 matches should be collected in each interval to make a valid average. After-hours travel time between 12am and 5am should be treated separately due to low volume.

There are next-step actions that could be taken to continue the analysis of this report:

• The Northwest Region ITS Software group at WSDOT is currently working on making an open socket connection to the cameras from a central server as described in this report. An averaging program is also being developed to smooth the data and output travel times. These processes were not completed in time to be included in the scope of this report and need analysis in the future. The suggested parameters such as interval (4 minutes) and minimum match count (10) should be implemented and tested. These parameters were based on the performance of the ideal segment (Segment 3). How do these numbers change for the other segments?

• A closer look at signal timing and reported travel times should be examined. How much does the current signal timing affect travel times? Can we put a quantifiable measure on how much the coordinated timing improves the corridor using travel time information?

• Given the analysis of this corridor, ALPR technology would be ideal for freeway travel time collection where there is limited access. How many target lanes in one direction would be needed for a reasonable resolution on a 6-lane freeway? Will high speeds have a negative impact on reader performance?

• How should the travel time data be presented graphically on the user end? What is the most useful method for engineers and/or the public?

REFERENCES

Mannering, Fred L, and Walter P. Kilareski. Principles of Highway Engineering and Traffic Analysis. New York: John Wiley & Sons, Inc, 1998.

Montgomery, Douglas C. Introduction to Statistical Quality Control. New York: John Wiley & Sons, Inc, 1996.

PIPS Techonology. U.S. Home Page. 08 June 2007. .

The Washington State Department of Transportation. Home Page. 08 June 2007 .

ACKNOWLEDGEMENTS

I’d like to thank the following people for their leadership and/or assistance to make this research possible.

Prof. Yinhai Wang Research Advisor

Mark Leth WSDOT Northwest Region Traffic Engineer

Sean Brackett WSDOT Northwest Region Signal Operations

Christian Cheney WSDOT Northwest Region ITS Software

Dongho Chang City of Everett Traffic Engineer

APPENDIX A

[pic]

[pic]

[pic]

[pic]

APPENDIX B

[pic]

[pic]

[pic]

................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download