Predictive Tracking of Continuous Object Boundaries Using Sparse Local Estimates

Environmental hazards (wildfires, floods, oil spills) are often modeled as “continuous objects” which evolve in space and time taking irregular shapes. Tracking their boundaries and accurately predicting their spatiotemporal spreading patterns is of paramount importance to combat their often catastrophic consequences. Wireless Sensor Networks (WSN) have been very instrumental for this purpose. However, current WSN-based methods require a prohibitively large density of deployed sensors to achieve a reasonable boundary reconstruction accuracy because they are based on the explicit identification of nodes close to the boundary. We present a novel approach that can track and predict the global boundary using only a relatively small number of distributed local front estimates. Our approach first filters and fuses the available sparse set of local front estimates (e.g. vectors of local orientation, direction and speed) and then uses the resulting information to reconstruct a smooth curve prediction of the evolving object's boundary at a future time. Moreover, since the uncertainty of the local front parameters is modeled, it can provide a heatmap representation of the evolving object, indicating the probability for each point in the area of interest to be reached at a future time by the spreading hazard. These predictive modeling capabilities when combined enable effective decision support for crisis management. We demonstrate that different types of continuous objects can be tracked with accuracy, even when only a relatively small number of noisy local front estimates is available. Our approach is practical since it can be applied in many situations where global boundary prediction updates are important to obtain by new sparsely distributed noisy local front estimates as soon as they reach a control center while the hazard is still progressing.