The updating of road network databases is crucial to many Geographic Information System (GIS) applications like navigation, urban planning, as well as emergency and disaster management. The development of a robust methodology for automatic or semi-automatic road extraction and change detection as well as “discovery of paths” (in military theater, desert areas etc.) is essential. Such a methodology has to provide accurate and up-to-date results albeit using noisy and infrequent sensor data. In this paper a new approach for a semi-automatic road extraction is presented that utilizes and combines output from particle filtering tracking (sequential Monte Carlo) and GIS techniques.
The target tracks may be generated from different sensor sources. In this paper we consider airborne platforms with GMTI (Ground Moving Target Indication) radar, which is able to continuously monitor the traffic in large areas day and night and at difficult weather conditions. The radar measurements however suffer from limited resolution, measurement noise, false alarms, and missed detections due to small target velocity or terrain shadowing. It is, therefore, difficult to extract the exact path of vehicles, and hence the road coordinates, directly from the sensor measurements. As an intermediate step we utilize particle filters to generate target tracks from the time series of sensor measurements.
The technique is applied to simulated road traffic in different scenarios such as straight roads, cross sections, and traffic roundabouts. The time series of the particle cloud is the input for the road extraction algorithm, which constitutes of different kinds of GIS methods. We apply Kernel Density Estimation (KDE) to generate a raster dataset where each cell has a weight according to the density of particles. Based on a threshold, a binary raster is created and converted in to a polygon shape file representing the road segments. Smoothing techniques are applied to provide a more realistic shape of the extracted roads. If available, a Digital Elevation Model can be used to exclude particle positions through raster calculations, in areas with large slopes that are unfeasible for road targets.
The comparison of the extracted roads with real maps proves successful semi-automatic road extraction in different scenarios. Examples are given in Figure 1. In combination with existing road maps or other landmarks, the coordinates of the extracted roads also can be used to calibrate the sensors and correct possible sensor bias.