Scalable Decentralized Partial State Estimation with Sensor Uncertainties Using Factorized Data Fusion

Decentralized target tracking and geolocalization problems are very important for a wide range of autonomous UAV applications, e.g. wildlife monitoring; security/suveillance; signal source localization; and wilderness search and rescue, just to name a few. It is often highly desirable for multiple UAVs to combine information to obtain better target state estimates. Ideally, in order to extract maximum information and maintain global consistency across all platforms, all UAVs would send observations from raw sensor data back to ground station or other centralized processing point for fusion. However, this approach scales poorly for large networks and in large environments, where communication quality cannot be guaranteed. This approach also becomes infeasible whenver sensor data is too expensive to communicate (e.g. image/video data). One alternative is to consider a parallelized/distributed approach such as decentralized data fusion (DDF), in which all UAVs combine local state estimates derived from Bayesian filtering using peer-to-peer message passing. It can be shown that DDF provies estimation results that are mathematically equivalent to centralized data fusion and has been successfully demonstrated for a number of UAV-based tracking and localization applications, e.g. see refs. 7–9. However, several important challenges remain, which prevent real-world UAV networks from achieving their full potential for cooperative target localization. Firstly, although target state estimation errors are tightly coupled to UAV state estimation errors (i.e. uncertainties in pose/attitude estimate), most tracking approaches assume that UAV/ownship uncertainties are neglible. However, precise and reliable geolocalization/tracking realistically requires accounting for ownship uncertainties in the fusion process. These uncertainties could, for instance, be induced by unknown sensor biases or parameters, e.g. for IMUs, gimbal-mounted cameras, or directional antennas. Proper accounting of ownship state uncertainties quickly becomes intractable and unscalable for decentralized estimation techniques, since these would require every vehicle in the network to maintain an estimate of the entire joint target and vehicle network state. This would not only be computationally expensive for onboard processing, but would also lead to excessive vehicle-to-vehicle communication costs that scale poorly with the number of vehicles and/or targets. Secondly, optimal DDF algorithms require vehicle networks to be connected in loop-free/tree topologies to avoid ‘data incest’ (i.e. improper ‘double counting’ of common information, whereby information that was already fused at some point in the tracking network is treated again as new independent information). This restriction can be overcome by using ‘conservative’ DDF techniques such as covariance intersection (CI) or the Weighted Exponential Product (WEP) fusion rule, which permit ad hoc communication topologies and are guaranteed to avoid data incest. However, these techniques can be overly conservative and can sacrifice too much newly gained information to maintain statistical consistency. This is highly problematic in the context of decentralized tracking with ownship uncertainties, since this can force vehicles to negotiate possible trade offs in target state information gain/loss at the expense of local ownship state information loss/gain. Ideally, the decentralized fusion process should ensure that all vehicles can keep losses in both target state and local ownship state information to a mutually beneficial minimum. In this work, we propose a novel scalable approach to conservative DDF for target localization with ownship uncertainties. Our approach builds upon the recently developed idea of factorized distributed data ∗Assistant Professor, Aerospace Engineering Sciences, University of Colorado Boulder, Boulder, CO, AIAA Member †Senior Researcher, Draper Laboratory, Cambridge, MA, AIAA Member ‡Graduate Research Assistant, Aerospace Engineering Sciences, University of Colorado Boulder, Boulder, CO, AIAA Student Member §Associate Professor, Aerospace Engineering Sciences, University of Colorado Boulder, Boulder, CO, AIAA Member