Automatic Compilation of 3d Road Features Using Lidar and Multi-spectral Source Data

While many commercial cartographic feature extraction systems process panchromatic and color imagery, few systems fully integrate airborne LIDAR data as a component of the feature extraction process. Most computational techniques seek elevation discontinuities at object boundaries or use the physics of the LIDAR signal to detect regions indicative of sharp edges and/or foliage. Automated road extraction presents a particular problem for these techniques because the height of the road surface is not significantly different from the height of the surrounding terrain. Our approach is to use co-registered airborne LIDAR and multi-spectral source (MSS) data as additional evidence for evaluation and extrapolation of road surfaces and centerlines within a commercial road network extraction system. Thus, road tracking provides the local context for source data processing in a relatively narrow search area by using LIDAR/MSS data to contribute to feature detection and delineation. LIDAR can provide the system with a direct determination of road height, pitch, and slope, while MSS can be used to estimate surface material properties. This work builds on previously reported work (Harvey, et al. 2006) that used multiple panchromatic image views to derive road elevation and slope while performing feature extraction. Incorporating co-registered LIDAR permits direct elevation determination and is less sensitive to occlusions and multi-image matching errors. In this paper we describe results in detection and delineation of road networks using co-registered LIDAR/MSS imagery. In addition to further automating the extraction process, we show that this approach improves road network accuracy and completeness, especially in complex urban environments. MOTIVATIONS AND BACKGROUND While most commercial cartographic feature extraction systems work with panchromatic (black and white) and color imagery, few systems fully integrate LIDAR data as a component of the feature extraction process. Commercial LIDAR analysis toolkits primarily are used to establish a digital elevation model (DEM) from the digital terrain model (DTM) which results from the LIDAR acquisition. During the DTM to DEM process, various man-made and natural features such as buildings and tree canopies can be "segmented" from the overall DTM since they exhibit sharp height discontinuities above the ground. These segments are represented as polygonal outlines of the boundary between the ground and the elevated features. Computational techniques that look for these height discontinuities at the object boundary, or use the physics of the LIDAR signal to detect "multi-bounce" indicative of foliage penetration, vary in effectiveness and depend on the complexity of the underlying terrain, the height of the structures being searched for, and the spatial resolution of the LIDAR compared to the size of the man-made structures. Automated road extraction presents a particular problem for these techniques since the height of the surface of the road is generally not much different than the height of the surface of the surrounding terrain. This is especially true in complex urban areas, which are among the most challenging locations for automatic cartographic feature extraction systems. With the exception of elevated highways, causeways, and bridges, the nature of road construction and civil engineering dictates that there exists a relatively wide area on the sides of roads without large, continuous, height discontinuities. Thus, in order to exploit the height information that is available in LIDAR data we have chosen to process the reflective surface which is co-registered with the LIDAR height data. ASPRS 2008 Annual Conference Portland, Oregon ♦ April 28 May 2, 2008 Likewise, there are many commercial MSI-HSI classification systems that can be used to search for materials which may be indicative of road construction. However, in many areas of the world, roads (cart tracks) are essentially unpaved paths with subtle visible texture changes, nearly indistinguishable in terms of surface material from the surrounding area. Per-pixel classifiers have the limitation that collections of similarly classified image points need to be aggregated into vector data suitable for M&S data repositories, including the generation of topology and functional attribution. These observations lead directly to our approach to use a feature-based system to guide the interpretation of LIDAR/MSS data. There are several recent results that employ airborne LIDAR data for road feature extraction. Alharthy and Bethel (2003) classify the LIDAR data using the signal intensity and height information, filtering the data to remove features unrelated to road objects. This work was focused primarily on urban road extraction where roads are delimited by significant depth discontinuities. Vosselman (2003) uses cadastral maps as a context within which to infer urban road regions from LIDAR point clouds. Filtering and fitting the point cloud data is a major emphasis in this work. Zhu, et al., (2004) detect road objects in high resolution color image data, then use LIDAR data to identify and connect roads across shadow regions. This work was also performed in the context of urban road extraction. Hu, et al., (2004) use a Hough transform to detect and extract gridded road networks from LIDAR data. Color imagery is used to classify and exclude vegetated regions from consideration in the network construction process. Finally, Clode et al. (2004) perform a classification of the LIDAR data by first using it to create a DTM, then identifying road regions by filtering based on intensity and local point density, as well as by comparing to the DTM. The road network is created from this result using connected components. As described in the following section, our road tracking process performs local analysis and maintains a dynamically updated history of road structure and appearance in panchromatic or color imagery. As a result, the use of LIDAR with correlated EO data becomes an additional source of information useful in the decision process in the road network control layer. This focus can be used to remove the need for assumptions about the road network structure or height discontinuities with the surroundings.