Photographic Data Extraction Feasibility and Pilot Study in Support of Roadside Safety and Roadway Departure Research

CHAPTER 2: RESULTS OF THE PHOTOGRAPHIC DATA EXTRACTION FEASIBILITY STUDY

Geographical coordinates found in FARS are a valuable resource to the roadside safety researcher. The limited roadway environment data were able to be assessed using panoramic view software, which provided an image of the crash scene according to geographical coordinates. Although the crash scene images were temporally discontinuous, they provided insight into the type and placement of the barrier cited in the crash data.

Traditional data acquisition has been inadequate to address the needs of FHWA. The photographic analysis sought to enhance the few relevant coded variables and limited attributes for those variables. However, issues exist in photographic analysis, such as unknown crash locations and inconsistent levels of detail, which require the activation of several operative assumptions.

CANDIDATE DATASETS

During the course of the pilot study, the researchers considered all relevant NHTSA datasets for supplementary data acquisition. FARS provided secondary access to crash scene images. The geographical coordinates were supplied to third-party panoramic view software, yielding images that were temporally discontinuous with the crash. NASS CDS provided near-crash time photographs, while NMVCCS provided on-scene crash photography.

The research team summarized the process for data filtering, photographic capture, and possible data codification parameters in table 2. The principal difference in filtering rested with the FHWA roadway departure definition applicable only to FARS. Owing to privacy protocol observed in NASS CDS and NMVCCS, variables and attributes necessary to apply the roadway departure definition were not reported. As a surrogate for the proper roadway departure definition, impacts occurring in the first event involving guardrails, concrete barriers, and impact attenuators were accepted as quasi-roadway departure crashes. FARS required a two-step process for extracting the geographical coordinates from the data files and then inputting the coordinates into panoramic view software to observe the crash location during some point in time not coinciding with the crash. NMVCCS provided the photographs linking the vehicle to the crash scene and final disposition, suggesting trajectory after impacting the roadside element. Finally, NASS CDS provided crash location photographs within days or weeks of the crash as well as post-crash vehicle conditions for vehicles accessible at the tow yard.

OPERATIVE ASSUMPTIONS

For the initial review of FARS images, the research team adopted the following assumptions to facilitate image review:

• If the image pre-dated the crash, the barrier would be the one found at the time of the crash.
  
  
• If the image was taken after the crash, the barrier would have been replaced with a similar barrier.
  
  
• In obvious cases of miscoded geographical coordinates, the following hierarchy would be applied to ascertain the crash location:
  
  
  1. Review of all vehicle crash events by scrolling along roadway segments.
     
     
  2. Consult street, cross street, and mile post information, especially in the case of elevation.

EXAMPLES OF CANDIDATE CRASH SELECTION FINDINGS

The research team filtered FARS, NASS CDS, and NMVCCS to compile a roadway departure dataset. They also entered the geographical coordinates into panoramic view software while using crash scene photographs as geographical coordinate surrogates for the datasets governed by privacy protocols.

FARS Dataset

With the publication of geographical coordinates in 2008 and retroactive provision for prior crash years by NHTSA, possible crash trajectories could be conceived. Starting in 2004, FARS started to disaggregate the crash into vehicle event units, which was previously only done in NASS CDS.^(6,11) With the coded vehicle events and the crash locations, it became possible for researchers to envision at least one crash trajectory, whereas previously, the crash scene was unknown to data analysts.

Figure 1 provides one possibility with respect to crash progression. From similar images, it became possible for practitioners and researchers to have a discussion relevant to roadway vernacular. For example, a police accident report may cite the involvement of a ditch, but a roadway designer may have posited the presence of an embankment. Additionally, questions with the coded data existed regarding the application of a jackknife condition to a pickup truck. Unknowns also existed, with respect to the precise location of impact and Terhune scale rollover classification, but the general environment in which the vehicle was operating was better described.

Original image: ©Google Earth®. (See Acknowledgements section for trajectory overlay.) Figure 1. Illustration. FARS data acquisition example showing possible vehicle trajectory.(2,12)

NMVCCS Dataset

Although NHTSA did not provide geographical coordinates in the NMVCCS dataset, the on-scene photography generally superseded the need for such flexibility, as seen in figure 2. Generally, from using one photograph, researchers were able to extract information relevant to the vehicle final disposition, vehicle damage, and roadside element damage. From this information, it was possible to postulate the path and interactions leading to post-crash vehicle final resting position. As the photographs were taken on scene, the roadside damage, if present, was readily evident and could be linked to specific coded elements like vehicle paint color and paint transfers to the roadside element. An additional boon to the study was the presence of measurement rods, which were used previously in NASS CDS to quantify vehicle damage. In NMVCCS, the measurement rods were used to highlight elements in photographs and unintentionally provided the barrier height, damage height, and damage width. This emphasis had a secondary effect of providing much needed roadside element measurement details. As this was not mandated in the NMVCCS photography protocols, these measurements were available inconsistently throughout the dataset. The other available attribute was the damage to the barrier and its components, seen visually and generally created by those unaware of roadside design precepts

Figure 2. Illustration. NMVCCS photographic data acquisition example.(1,4)

NASS CDS Dataset

NASS CDS data were the final dataset considered.^(6,11) The photographic image capture had the unintended benefit of improvement after the publication of the NMVCCS data. NASS CDS images before NMVCCS focused on the vehicle and the occupant interaction with the vehicle, but after the on-scene NMVCCS experience, the crash environment images improved markedly from an FHWA perspective. Although the photographs were taken within days or weeks of the crash, a benefit existed in knowing the precise crash location. Based on the vehicle events recorded, the crash location and vehicle damage photographs might provide information relevant to possible trajectory.

NASS CDS photographic analysis required a two-step process to understand the crash scene and infrastructure as well as the vehicle damage and its possible interaction with the roadside environment. Most vehicles carried quantitative damage measurements conforming to Society of Automotive Engineering standards. These were also the same type of measurement rods that were repurposed in the NMVCCS dataset.

Figure 3 provides a summary of data relevant to the roadside element study extracted from NASS CDS Case 2007-09-013.⁽⁶⁾ The left image provides the roadside element, the deformed metal guardrail. The vehicle damage can be discerned from the center and right images. In this case, the front of the vehicle made contact with the barrier, followed by a subsequent impact with the barrier, culminating in a one-quarter turn driver-side rollover. The rollover was identified as a climb-over probably associated with the guardrail damage shown in the left image. In this case, the vehicle scaled the barrier, causing the vehicle to roll to the opposite side of the barrier.

Figure 3. Illustration. NASS CDS data acquisition example showing damaged roadside elements (left, W-beam strong post) and vehicle damage (center, back/left oblique and right, front with measurement rods).(6)

FEASIBILITY STUDY DATASET SELECTION

FARS was the first dataset considered because the availability of geographical coordinates allowed for a wide range of panoramic view image options.

FHWA has been a limited user of the NHTSA dataset, especially in problem identification, as in the case of the roadway departure data definition.⁽⁵⁾ Details relevant to developing crash testing protocols and roadside design parameters have been unavailable. It is believed that at minimum, the photographic data might provide valuable modeling inputs associated with different crash types. The supplemental data acquisition was developed based on the initial inventory style identification through FARS and was refined to include roadside element damage, vehicle damage, and vehicle trajectory.

For purposes of this current study, FARS and NMVCCS crashes were selected for consideration from a pool of crashes filtered using the FHWA roadway departure definition for FARS and the first event roadside element impact surrogate for NMVCCS. The images were reviewed per table 3. The case review process was outlined in Leveraging New Technologies to Capture Infrastructure Data to Supplement Roadside Crash Analyses.⁽¹³⁾

Table 3. FARS crash panoramic view image classification.

The FARS photographs fell into five major classification rubrics, translating into the photographic classification (see table 3). Upon reviewing the NMVCCS crashes, these were further simplified into a color usability classification system (green, yellow, and red).

The FARS green category included the generally useable classification, the yellow category contained the possibly useable classification, and the red category comprised the remaining rubrics (see table 4). Additionally, green was merited if the coded roadway element was confirmed by crash scene image, with at least one scenario emerging for modeling. Yellow was assigned if the coded roadway element was confirmed at proximate coordinates or confirmed by traffic way identifiers permitting the update of a coded roadway element, thereby improving the roadside element description based on coded location information or suggesting at least one scenario for modeling. Finally, red was assigned if geographical coordinates were unpublished, images were overexposed, or the street view was unavailable, as in the case of international border crossings.

Table 4. FARS candidate crashes by frequency and usability.

The same categories were applied when assessing NMVCCS crashes (see table 5). It was found that 71 percent of the first event roadway departure FARS crashes were classified as green and yellow collectively (table 4), and 100 percent of the NMVCCS crashes were classified as green (table 5). The FARS green and yellow aggregation and the NMVCCS green crashes were deemed suitable for consideration in the photographic data extraction pilot study as candidate results for the supplementary dataset.

Table 5. NMVCCS candidate crashes by frequency and usability.

FINDINGS OF THE FEASIBILITY STUDY

The goal of this study was to confirm the presence of coded roadside elements and determine their placement. It provided greater detail for the limited coded attributes, including uncoded roadside element type, and related them to the coded crash events. Finally, it provided insight into uncoded data with external images with full panoramic scene views for crashes with geographical location as well as NHTSA images with damage detail as a surrogate for datasets disallowing the provision of precise crash scene locations. These findings suggested the efficacy of the feasibility study and warranted continuation in the form of a pilot study.