Data and methods
Overview
This study investigates mobile LiDAR's dense ground-level collection capabilities to assist the urban design/analysis and transportation research community with two specific objectives.
The first objective is to provide descriptive statistic matrices that may be used as alternatives to some of the audit-based methods for streetscape evaluation with more precise and accurate data. Since the research strand tends to revolve around the walkability aspects of urban streetscapes, we attempt to focus on the sidewalk areas of busy downtown-like contexts. For these streetscapes, we also evaluate mobile LiDAR's ability to collect vertical building-face data that aerial LiDAR misses such as window proportions, awnings, and commercial signs.
Our second objective is to determine mobile LiDAR's capability of fitting small streetscape such as street signs, traffic lights, and streetscape furniture with a precise, voxel matrix. Mobile LiDAR should be highly effective for this level of analysis because our voxel size can be refined more than the 5-ft vertical intervals used in the previous research (Golombek and Marshall 2019, 2020).
Data/data collection
A Trimble MX9 was used to collect data for this study. The MX9 contains a spherical imaging system and three oblique view cameras. The MX9 contains Riegl VUX-1HA dual laser scanners that measure at 1000 kHz with the scanner mirror rotating at 250 revolutions per second and a field of view of 360°. A Real-Time-Kinematic (RTK) base station was set up for post-processing purposes on a nearby National Geodic Survey (NGS) monument set at a 1-s interval collection rate. To confirm system accuracy, 13 ground control points (GCP) were established throughout the survey area that were no more than a half-mile apart from each other. Riegl's collection platform uses an internal 'multiple-time-around' procedure, which eliminates range ambiguities and helps remove data noise.
A four-mile segment was surveyed along Montview Boulevard in Denver, Colorado, between Colorado Boulevard and Havana Street. An additional nearby mile of roadway through a mixed-use town center area along E. 29th Avenue in Denver was also collected. Each of these segments was traveled twice, once in each direction. Since many features of interest are concentrated on the sides of streets, doing so provides additional assurance for full collection of features.
POSPac ® was used to process telemetry of the Global Navigation Satellite System (GNSS) system, and we used RTK base station data to compute a corrected location of the vehicle path. Trimble Business Center (TBC) software was used to register/adjust overlapping data from the multiple runs. TBC also computed point cloud colorization using the registered images that were collected in-sync with the scans.
The combined runs were exported into LAS 1.4 format and 100-ft by 100-ft tiles. A control report was run against the GCPs that returned a root mean square error (RMSE) of 0.06-ft/1.83-cm RMSE horizontal and a 0.122-ft/3.72-cm RMSE vertical. Point cloud density consistently averaged over 2000 points per square meter. Target range was well over 100m, which was well above the tallest feature in the streetscapes.
We also obtained parcel data from the City and County of Denver to determine the right-of-way (ROW) for the Montview segment, as discussed below.
Methods
Creating streetscape sample areas
Similar to previous studies on quantitative streetscape measuring (Golombek and Marshall 2019, 2020), we seek to develop quantitative statistics of streetscape features. This first required defining streetscape corridor segments. Recent streetscape mapping studies used a method to divide the pre-determined segments into Thiessen proximal polygons (Golombek and Marshall 2019, 2020). To create a Thiessen polygon layer, only a point feature class is required, and the output Thiessen polygon dataset is configured where each individual non-overlapping polygon is closest to its associated input point. In the Montview dataset, we attempted to avoid segment cutoff at intersections since urban intersections tend to be focal areas with high activity and many traffic-related features we prefer to keep together. By establishing the input point feature class with these intersection centroids, the segment breaks occur at a subtle location somewhere between these intersections. Figure 2 depicts the Thiessen-based streetscape segments we created for Montview Boulevard using the ESRI ArcGIS® Thiessen polygon-processing tool. The town-center streetscape polygons were created manually since the area of interest contained only a few small blocks.
Additionally for the Montview section, we designed the streetscape sample areas to include the full street right-of-way (ROW) that extend beyond the street curb. Extending the streetscape to include these peripheral views is important since the majority of objects in urban roadways that fall within a road user's peripheral view are located off the street itself. Therefore, we utilized Denver's parcel data around the area of interest (AOI), merged the parcel data together, and used an ESRI ArcGIS® erase function that erased all non-streetscape ROW areas from our AOI. For the Montview section, we created six streetscape polygon segments utilizing this method.
For the town center area, our approach was a little different. Since the focus is more on streetscape features related to walkability, we simply digitized the areas between the curb and the building, enclosed with a crosswalk on each side. We identified six sample streetscape areas, as shown on Fig. 2, that we assess in this study.
Creating streetscape voxels for positional feature analysis
A key objective of this study is to incorporate voxels to quantity streetscape features and compile descriptive statistics. Breaking up a streetscape into voxels creates user-specified height zones. A voxel is a pixel with a height or third-dimension component that can be useful for analyzing LiDAR data. Similar voxel methods are utilized in our previous street-feature voxel analysis studies (Golombek and Marshall 2019, 2020). In those papers, we determined that—given aerial LiDAR point densities—a 3-ft by 3-ft horizontal and a 5-ft vertical setting was optimal and appropriate.
The accuracy and density of mobile LiDAR data, however, allows more flexibility in setting voxel parameters. Lehtomaki et al. used a fix voxel size of 20 cm3 for their street feature study. Yang et al. used a super-voxel approach where the voxel size was more focused on the feature itself, though their voxel size always seemed greater than 0.5 m3. With over a thousand points per meter collected and utilizing US feet as the linear unit, we suggest setting each voxel to a single cubic foot, or 1-ft by 1-ft by 1-ft (x, y, z) dimension. We suggest this parameter because this dimension facilitates a simple calculating dimension and solid measuring basis for descriptive statistic metrics needed in the town center area. Ewing and Clemente often reference measuring features from ground to certain levels, like eye-level visibility (Ewing and Clemente 2013). If a streetscape was, for example, 500-ft long from crosswalk to crosswalk, 50-ft wide from storefront/building to curb, and 15-ft tall from ground to the above awnings, that streetscape quantitatively would total 375,000 cubic ft. If we calculate that 10 landscape trees fill 7500 single cubic-foot voxels, then we can estimate that trees comprise 2 percent of that streetscape.
LiDAR feature extraction methodology/classifying the data
For the four-mile Montview segment, the goal is to devise comprehensive descriptive statistics of streetscape features that can be applied to various urban transportation research studies. Other than for street sign extraction, the existing literature did not provide viable automated approaches. Even the street sign extraction methods yielded results well below 100% (with a high degree of false positives and false negatives) and have not been thoroughly tested at a macro scale. Gargoum et al. developed highly accurate automated street sign extraction results (Gargoum et al. 2018), though testing was primarily in rural areas where signs are isolated. Even if we utilize published automated methods, we would still have to manually assess the entire streetscape to check for false positives and false negatives. We also note, after previewing the Montview section, that many street signs are on the same horizontal poles as traffic lights, and some are small and/or obscured. For example in Fig. 3, a 'No Parking' sign may be on a light pole or power pole, but if it is not wider than the pole, the current automated methods would likely miss these signs. Additionally, the streetscape furniture and traffic lights require manual classification since we did not find adequate methods to automate extracting those features. An accuracy assessment on manually classified data is not necessary since many automated LiDAR classification experiments, including ones listed above, utilize manual classification for their control to assess accuracy. The accuracy specifications of the data are mentioned with the data collection platform specifications above.
Since our four-mile sample area is not overwhelmingly large, we utilized a manual classification approach with some automated LiDAR functions. The data's control reports, noted above, confirmed very high vertical and horizontal point-cloud accuracy. Also, automated LiDAR methods are often tested against manual classification for control, and we saw the same with automated urban mobile LiDAR feature extraction (Yang et al. 2015). We auto-filtered the data for ground points and also used an automated process to classify the remaining above-ground features together on a single LiDAR classification layer. We utilized Merrick & Company's MARS® LiDAR processing software, which has a Google Earth simultaneous location feature, to follow the streetscape point-cloud in MARS®. Paired with Google Earth on a separate screen, this allowed us to visually identify all street signs, traffic signals, and street furniture features (Fig. 4).
Once completed, we utilized a voxel application tool in MARS® that incorporates ground level and builds a user-specified voxel grid set to 1-ft by 1-ft by 1-ft (x, y, z) dimensions. This tool then applies the selected classified street furniture and street sign features into the voxel grids and exports a .dat Envi raster dataset, with each band within the .dat file representing a unique user-specified voxel height zone. Each grid cell that is populated also has a count value that specifies how many LiDAR points fell into that 3D voxel grid cell. For this study, we are not very concerned with the amount of data in each grid cell, but rather the simple fact that classified LAS points are inside that grid cell. Table 1 presents a sample descriptive statistics output.
For the town center area, our descriptive statistic zones for the walkable streets are different because we focus more on the walkable areas between the curb and the buildings. For this, we use a different approach because we are interested in collecting and measuring features affiliated with the buildings themselves such as the windows, awnings, and hanging signs. Our approach for collecting these objects differs because flat objects against buildings, such as windows, cannot be extracted like typical streetscape objects. Also, LiDAR often gets obstructed at windows and does not yield returns.
We extracted windows and awnings with mobile LiDAR with the help of a web-based solution by Orbit GT technologies. The Trimble MX9 platform collected panoramic images in conjunction with the LiDAR. Since the frames of the windows are well defined based on the void of LiDAR points within the frames, the Orbit LiDAR software has an interface that allows 3D feature tracing of the LiDAR data. For windows and awnings, we essentially created vertical polygons. After the windows and awnings were drawn in the 3D Orbit software interface, it exported the polygons into a Google KML file, which were then converted into a 3D shapefiles. Since the features are in 3D, calculating the area of these polygons must be captured in 3D as well. For this, we created an automated model in ESRI's® Model Builder to iterate through each polygon feature and created a temporary Triangulated Irregular Network (TIN) for each window or awning polygon. We then ran surface information statistics on each polygon, utilizing the TIN as the surface, which enables the area of each polygon to be calculated. This process created area statistics for features like windows that can detected with LiDAR but not directly extracted and classified. As mentioned above, urban design researchers such as Ewing et al. take into account quantity of features at eye level for perceptual qualities. Utilizing data in the LiDAR viewer allows an analyst to calculate 2D face area of a building (from ground to the awning), and after the TIN process for windows is complete, area and proportion of windows were calculated.