Algorithms to Convert Basic Safety Messages into Traffic Measures (2022)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

15   Data for Developing Algorithms to Convert BSMs into Traffic Measures This chapter provides a summary of the data used to develop, test, and validate the seven measures of interest to the NCHRP 03-137 Project Panel. This chapter also provides an overview of the Trajectory Conversion Algorithm (TCA) used to emulate CV BSMs from the simulated vehicle trajectories. The project team used synthetic BSMs generated using the TCA from simulated vehicle tra- jectories. Section 2.1 provides a description of the data used for the study and a discussion of ground truth calculation. Section 2.2 provides a description of the approach for generating synthetic BSMs. 2.1 Data Used for Algorithm Development, Testing, and Validation The project team used simulated data to develop, test, and validate the algorithms. Table 4 presents an overview of the calibrated traffic simulation models that the team has access to. The simulated trajectories were used to estimate the ground truth as well as to generate synthetic BSMs using the TCA (Section 2.2). Urban Freeway/Dense Grid: New Jersey Turnpike/Jersey City serves as the urban freeway and dense grid network example. A PARAMICS model was built and calibrated by the NYU team (New York University UrbanMITS Laboratory and Rutgers RIME Laboratory 2016, Bartin et al. 2018). The PARAMICS simulation model is a 3-h microscopic simulation model of mainline New Jersey Turnpike (NJTPK) and downtown Jersey City with additional key road- ways and crossings in New Jersey and Manhattan calibrated for AM peak period. The model provides details of the toll plaza operations with a toll plaza lane selection algorithm developed by the team and comprehensive road network approaching the Holland Tunnel. The model includes an incident that took place on the Newark Bay-Hudson County Extension (NB-HCE) of the NJTPK milepost (MP) 7.0 on the left lane for 35 min and resulted in a closure of the left lane with a vehicle passing speed of 15 mph. Note that these are the usual characteristics of commonly observed incidents at this location. Figure 1 shows the simulation network in PARAMICS and the shaded area where the incident was modeled. Urban Arterial: The Flatbush Avenue, NY, simulation model that uses SUMO open-source simulation environment was used as the urban arterial network. This model was developed and calibrated by the New York University (NYU) team as a part of the NYC Connected Vehicle Pilot Program (New York City Department of Transportation 2020). Specifically, a 1.6-mile road segment on Flatbush Avenue between Tillary Street and Grand Army Plaza in Brooklyn, New York City, is the study area. Flatbush Avenue is a bidirectional, North–South urban corridor C H A P T E R 2

16 Algorithms to Convert Basic Safety Messages into Traffic Measures Simulated Network Source Simulation Software/ Calibration Status Type of Network Network Geometry Traffic Conditions New Jersey Turnpike/ Jersey City New York University UrbanMITS Laboratory and Rutgers RIME Laboratory (2016) PARAMICS/Calibrated to AM peak average + traffic incident conditions Urban freeway and dense grid NJTPK Newark Bay- Hudson County Extension (NB-HCE): Exit 14 to Holland Tunnel 7 mi x 3 mi 93 intersections Moderate congestion (NJTPK) Moderate to high congestion (Jersey City) Flatbush Avenue, NY New York City Department of Transportation (2020) SUMO/Calibrated to AM peak Urban arterial corridor 1.6 mi x 0.7 mi 17 intersections Moderate to high congestion Accident count (2012– 2014): 1,128 injuries, 8 fatalities I-405 Corridor, Seattle, WA WSDOT and USDOT (2019) VISSIM/Calibrated to AM/PM peak for six operational conditions (work zone, rain, incidents, low visibility, severe bottlenecks, and low demand) Urban freeway corridor 29.5 miles long 2 bottleneck locations Severe congestion with frequent rain/fog Table 4. Calibrated simulation models used by the project team. Source: New York University UrbanMITS Laboratory and Rutgers RIME Laboratory 2016 Figure 1. Urban Freeway/Dense Urban Grid: New Jersey Turnpike/Jersey City.

Data for Developing Algorithms to Convert BSMs into Traffic Measures 17   with eight lanes, four in each direction, with a median from Tillary Street to Fulton Street and six lanes, three in each direction, from Fulton Street to Grand Army Plaza. There is one parking lane on each side. Intersections within two or three blocks from Flatbush Avenue & Tillary Street and Flatbush Avenue & Grand Army Plaza are also included as buffer intersections in the study area. Roadway geometry, lane usage, link capacity, speed limit, lane and turn connectivity, and other parameters were thoroughly updated to represent 2018 conditions. Besides network topology, other basic road network information, such as signal timings and bus stops, is obtained and implemented when developing the simulation network. The simulation network built using SUMO is shown in Figure 2. The simulation network is modeled for the morning peak period (between 6 a.m. and 10 a.m.) and is calibrated for both the operational measures (i.e., link volumes and travel time) and safety measures (conflicts identified by time to collision) by applying the simultaneous perturbation stochastic approximation (SPSA) approach (Yang 2012, Sha et al. 2020). See New York University C2SMART Center (2020) for more detailed information. Urban Freeway: The I-405 Corridor, which is a 29.5-mile-long major commuter corridor in the Seattle, WA, area, as shown in Figure 3, was used as the urban freeway network. The corridor is subject to periods of high travel demand and congestion and experiences consider- able travel time variability due to dynamic incident patterns and frequent rain and fog. The simulated data set was generated as part of a separate project that Noblis conducted for the FHWA’s Traffic Analysis Tools (TAT) Program (Wunderlich et al. 2019). The Washington State Department of Transportation (WSDOT) provided FHWA and Noblis with traffic, travel time, incident, and weather data for 2012. As part of the TAT project, Noblis first removed weekends and holidays from the data set, which resulted in 196 weekdays. Noblis clustered the days into six operational conditions for the study—low demand, low visibility, weather and incidents, many incidents, bottleneck trouble, and few incidents—and a representative day was selected for each cluster. WSDOT also provided a VISSIM model of the I-405 network as part of the TAT project. Using this VISSIM network as an initial base model, Noblis calibrated VISSIM models for each of the six operational conditions. The simulation period was for the AM peak period from 5:30 a.m. to 10:30 a.m. See TAT Volume III 2019 Update (Wunderlich et al. 2019) for more information on the clustering and calibration of the I-405 model. The simulated vehicle trajectories from the six calibrated VISSIM models were available for the NCHRP 03-137 research project. 2.1.1 Additional Validation Data For the validation of the algorithms for the three safety measures, the team used new emulated BSMs generated by the TCA from the Flatbush Avenue & Tillary Street urban intersection in Brooklyn, New York City, and a segment of FDR urban freeway in Manhattan, New York City. Specifically, Flatbush Avenue & Tillary Street is one of the busiest intersections along the Flatbush Avenue corridor. In 2019, approximately 1.5 h of traffic surveillance video was recorded in the morning peak period (6 a.m. to 10 a.m.) across four weekdays. The project team had access to these data to conduct research for this project. Anonymous vehicle trajectories, including the longitude and the latitude of the center of each vehicle, as well as the speed and acceleration, were extracted from the recorded drone videos (New York University C2SMART Center 2019, Data From Sky 2020). The location of the Flatbush Avenue & Tillary Street inter- section, as well as the extracted vehicle trajectories taken by the drone’s camera, are shown in Figure 4. See New York University C2SMART Center (2019) for a detailed discussion. The sample size of the field-collected vehicle trajectory data used for algorithm validation is only

18 Algorithms to Convert Basic Safety Messages into Traffic Measures Source: New York University C2SMART Center 2020 Figure 2. Urban Arterial: Flatbush Avenue.

Data for Developing Algorithms to Convert BSMs into Traffic Measures 19   Source: Google Maps Figure 3. Urban Freeway: I-405 Geographic Network.

20 Algorithms to Convert Basic Safety Messages into Traffic Measures Source: Third chart from Yang et al. 2021 Figure 4. The location of the intersection and the extracted vehicle trajectories. Figure 5. FDR Urban Freeway segment. half of the 3-h simulated trajectory data used for algorithm development and test. This is a key limitation of these data. For the urban freeway scenario, an urban freeway segment of approximately 260 ft along the FDR in Manhattan was selected based on the best availability of the field data. This freeway segment is covered by the street-level traffic surveillance camera mounted at 79th Street, as shown in Figure 5. Approximately 4 h of traffic surveillance video in the morning peak period

Data for Developing Algorithms to Convert BSMs into Traffic Measures 21   (6 a.m. to 10 a.m.) were recorded on Thursday, May 2, 2019. The video was then processed to obtain anonymous vehicle trajectories. Camera distortions in the vehicle trajectories have been corrected. See New York University C2SMART Center (2019) for a more detailed discussion. Compared to the 1,900-ft length of the urban freeway used for algorithm development and test, the length of the freeway segment used for the validation is relatively short (260 ft). An additional data set—traffic crashes that occurred at these locations—was also used for safety algorithm validation. For both the urban intersection and urban freeway scenarios, motor vehicle collision data in 2019 were obtained from the NYC Open Data Portal (NYC Open Data Portal 2020), as videos were recorded in 2019. 2.1.2 Ground Truth Calculation Ground truth represents reality—what happened on the ground. Thus, ground truth can be accurately captured only if precise information is available from all vehicles all the time. Ground truth is not the same as 100% market penetration of CVs. The key difference between ground truth and 100% market penetration rate is that unless BSMs are “heard” by a device, they are lost. Second, the lane-specific precision problem is still under resolution. Performance measures generated by traffic simulation software should not be used as the ground truth because the calculations of these measures vary by simulation software. For example, queue length has several nuanced interpretations and implementations in leading traffic simulation tools. To ensure consistency, the research team developed or re-used algorithms to calculate the ground truth measures from the traffic simulation outputs that describe the vehicle dynamics (e.g., position, speed, and acceleration rates) of every vehicle in the network. Chapter 3 and Chapter 4 include approaches for ground truth calculation for each measure. 2.2 Approach to Generate Synthetic BSMs Using TCA This section discusses the high-level analytical process for generating synthetic BSMs from field or simulated vehicle trajectories for specific market penetrations of CVs (see Figure 6). In Figure 6, the upper left corner shows a time–space diagram of vehicles traversing a congested roadway Figure 6. Processing vehicle trajectories to synthetic BSMs to measures.

22 Algorithms to Convert Basic Safety Messages into Traffic Measures section; each line corresponds to a vehicle trajectory (combinations of vehicle speeds and loca- tions for a unique vehicle). These may be GPS traces from vehicles traveling on the roadway or vehicle trajectories generated by a traffic simulation model. Next, the research team ran- domly selected a subset of vehicle trajectories to represent the various market penetration rates of CV technology. Then, the extracted vehicle trajectories are fed into the TCA open-source tool. The TCA generates BSMs based on the SAE J2735 standard. These BSMs are then processed (see left bottom corner) and analyzed to generate traffic measures. The TCA is designed to test different strategies for producing, transmitting, and storing CV messages. The TCA uses vehicle tra- jectory data, roadside equipment (RSE) location information, cellular region information, and strategy information to emulate the messages that CVs would produce. These can be transmitted by either dedicated short-range communication (DSRC) or cellular communication. Users also have the option to adjust the market penetration of CVs, the message transmission frequency, and many other parameters. The TCA is available to download from the U.S. Department of Trans- portation (USDOT) ITS CodeHub. 2.3 Lessons Learned from Data Preparation and Fusion This section highlights a few lessons learned from this project with respect to data preparation and fusion. • Data from multiple sources should be transformed into a consistent frame of geo- graphical and temporal resolution. To successfully fuse data from multiple files, their units of measurement and types must match. Data elements such as distance (e.g., link lengths) and time often have varying units of measurement. For example, link length may be in feet or meters, and time may be in total seconds or in hours, minutes, and seconds. Additionally, data elements could be stored or read as different data types. For example, time could be stored as a numeric float value, datetime object, or string. Read the metadata carefully to check units of measurement and data types across data sets and make the data consistent across types, formats, and units before fusing or integrating disparate data sets. • Supporting data are necessary to contextualize BSMs. Although BSMs contain a plethora of detailed vehicle position and status information, this information remains abstract unless it is somehow connected to the roadway network. Fusing BSMs with other data is necessary to achieve this contextualization. Depending on the measure, supporting network files may contain geometric data on link coordinates, intersection stop line coordinates, the number of lanes per link, or operational insights such as signal timing plans and other network information necessary for measures estimation. • Limited GPS precision makes developing lane-level measures difficult. Synthetic BSMs from simulated vehicle trajectories were used in the study. Although synthetic BSMs fall neatly into roadway links and lanes, real-world BSMs often do not due to GPS positional errors. Several factors contribute to GPS positional errors, including persistent GPS drifts, signal interference, or reflection off tall buildings and other surfaces. These errors can result in a vehicle appearing to be on a different lane, segment, or altitude, which is especially prob- lematic for differentiating movement at a multi-level interchange. GPS positional errors make it difficult to develop lane-specific traffic measures using BSM data. Because the goal of this project is for agencies to be able to use these algorithms with their real-world BSMs rather than simulated BSMs, estimation of lane-level measures using uncorrected BSM data is too ambitious at this time. Sorting out GPS drift and other positional errors is likely best addressed using multiple independent sources of data rather than relying on any one source to represent “ground truth.”

Data for Developing Algorithms to Convert BSMs into Traffic Measures 23   • BSM aggregation (spatial and temporal) is necessary for mobility measures. BSM data are designed specifically for safety applications. The extremely granular BSM data with attributes of low latency and specificity (spatially and temporally) are essential for near-instantaneous detection of movements and behaviors that could lead to crashes. But it is challenging to use the highly detailed BSM data for estimation of mobility measures without removing redundancies, managing outliers, and smoothing the data. It may be necessary to engineer features that aggregate BSMs spatially and temporally to obtain a more comprehensive view of the traffic conditions for estimation of mobility measures.