Multi-modal identification and tracking of vehicles in partially observed environments

High-quality positioning is of fundamental importance for an increasing variety of advanced driver assistance systems. GNSS-based systems are predominant outdoors but usually fail in enclosed areas where a direct line-of-sight to satellites is unavailable. For those scenarios, external infrastructure-based positioning systems are a promising alternative. However, external position detections have no identity information as they may belong to any object, i.e. they are anonymous. Moreover, the area covered by external sensors may contain gaps where objects cannot be observed leading to a correspondence problem between multiple detections and actual objects. We present a global tracking-by-identification approach as extension to existing local trackers that uses odometry sensor data of vehicles to find the corresponding subset of external detections. Thus, our approach enables the assignment of anonymous external detections to a specific vehicular endpoint and the estimation of its current position without requiring an initial location. The problem is decomposed resulting in a two step approach. The first algorithm determines possible track segment combinations which are used as track hypotheses. The track hypothesis generation algorithm considers spatio-temporal relationships between track segments, thus avoiding exponentially growing complexity inherent to data association problems. The second algorithm compares track hypotheses to the relative vehicle trajectory using pseudo-distance correlation metrics. In a detailed evaluation, we demonstrate that the proposed approach is able to reliably perform global tracking and identification of camera-observed vehicles in real-time, despite relatively large coverage gaps of the external sensors.