Context Exploitation in Data Fusion

Complex and dynamic environments constitute a challenge for existing tracking algorithms. For this reason, modern solutions are trying to utilize any available information which could help to constrain, improve or explain the measurements. So called Context Information (CI) is understood as information that surrounds an element of interest, whose knowledge may help understanding the (estimated) situation and also in reacting to that situation. However, context discovery and exploitation are still largely unexplored research topics. Until now, the context has been extensively exploited as a parameter in system and measurement models which led to the development of numerous approaches for the linear or non-linear constrained estimation and target tracking. More specifically, the spatial or static context is the most common source of the ambient information, i.e. features, utilized for recursive enhancement of the state variables either in the prediction or the measurement update of the filters. In the case of multiple model estimators, context can not only be related to the state but also to a certain mode of the filter. Common practice for multiple model scenarios is to represent states and context as a joint distribution of Gaussian mixtures. These approaches are commonly referred as the join tracking and classification. Alternatively, the usefulness of context was also demonstrated in aiding the measurement data association. Process of formulating a hypothesis, which assigns a particular measurement to the track, is traditionally governed by the empirical knowledge of the noise characteristics of sensors and operating environment, i.e. probability of detection, false alarm, clutter noise, which can be further enhanced by conditioning on context. We believe that interactions between the environment and the object could be classified into actions, activities and intents, and formed into structured graphs with contextual links translated into arcs. By learning the environment model we will be able to make prediction on the target’s future actions based on its past observation. Probability of target future action could be utilized in the fusion process to adjust tracker confidence on measurements. By incorporating contextual knowledge of the environment, in the form of a likelihood function, in the filter measurement update step, we have been able to reduce uncertainties of the tracking solution and improve the consistency of the track. The promising results demonstrate that the fusion of CI brings a significant performance improvement in comparison to the regular tracking approaches.