Hierarchical multi-hypothesis tracking with application to multi-scale sensor data

In the target tracking literature, the benefits of a centralized architecture are well documented. Similarly, the performance of decentralized estimation algorithms is often measured against the upper bound provided by centralized, full-information estimation. However, the benefits that exist in single-stage centralized tracking must be measured against the computational cost that is incurred. Under certain circumstances, even when sensor data is available at a centralized processing center, a decoupled approach to multi-sensor multi-target tracking may provide improved performance. In this paper, we propose a centralized, hierarchical framework for data fusion, which we refer to as hierarchical multi-hypothesis tracking (HMHT). The HMHT architecture requires multiple fusion modules that perform combined detection-level and track-level tracking, with all fusion modules having access to sensor data. We explore the benefits of the HMHT approach in multi-sensor settings where sensor update rates and the information content of the data differ significantly. The approach also shows promise for processing low-observability bearings-only data, as well as for road-constrained tracking.