Kernel Based Method for Distributed Derived Feature Tracking in High Dimensions

Modern sensing systems are increasingly heterogeneous and decentralized. These systems require new methods for efficiently combining data across distributed networks when centralized data fusion centers are impractical due to communications limitations. Consensus and innovation algorithms are a class of algorithms for fusing sensor data over distributed networks without the need for full connectivity to a centralized system. We present a novel method for combining a consensus and innovation framework with kernel density estimation to track complex non-observable features of targets over a high dimensional space. The goal of our method is to track multiple targets over time while categorizing their long term behavior. Instantaneous features of the targets are used both as tracking tools and combined over time to establish higherorder features of the targets' long term behavior. We assume that the communication bandwidth in the network is low, and that real-time identification of specific long term behaviors, such as a pattern of suspicious activity, is a priority. We compare the capabilities and limitations of our method with common modern tracking methods including particle filtering and multihypothesis testing. Results are given for an example scenario of a heterogeneous set of sensors identifying a suspicious target vehicle from traffic data. The instantaneous measured features include the location, color, speed, and fuel consumption.