On Tracking-driven Road Map Extraction from Gmti Radar Data

For analyzing dynamic scenarios characterized by many ground moving vehicles, airborne GMTI radar is a well-suited sensor due to its wide-area, all-weather, day/night, and real time capabilities (GMTI: Ground Moving Target Indicator). The generation of GMTI tracks from these data is the backbone for producing a “recognized ground picture” as well as for analyzing traffic flows. In this paper we dicuss the benefits of GMTI tracking in view of extracting road map information from GMTI data. The resulting tracking-generated road maps are highly up-to-date and fairly precise. Moreover, their accuracy is quantitatively described. The proposed approach to road map extraction is essentially based on a temporal integration of the received sensor data and by this differs in nature from methods based on pattern recognition in a single image. The underlying idea is rather simple: By definition, the track of a road moving target provides an approximation of the underlying road. As GMTI tracking is a highly challenging task, the quality of GMTI tracks, however, is often insufficient for road map production due to false returns, missing detections, Doppler blindness, fading phenomena, and other reasons. In this context, retrodiction techniques can provide significant improvements. Being a generalization of standard smoothing techniques to multiple hypothesis tracking (or, more generally speaking, Gaussian sum or particle filtering), retrodiction provides fairly precise estimates of the target kinematics at past time instants by exploiting the sensor information available up to the present time. From ‘retrodicted tracks’ the road position and related tangential vectors to the road can easily be derived. For calculating additional supporting vectors ‘continuous time retrodiction’ is proposed. We indicate how the generated road map information is exploited in the tracking loop resulting in even more precise tracks. A procedure for iteratively producing high-precision road maps due to changing sensor-to-target geometries is sketched. The proposed approach is illustrated be a simulated example providing hints to the achievable road map accuracies. An important application is the mitigation of sensor registration errors by matching the produced sensor individual road maps with each other and with geo-referenced road maps available in a topographical data base.