Modelling and Characterisation of Detection Models in WAMI for Handling Negative Information

Abstract

Abstract : The objective of the project reported here was to develop, implement and evaluate a model of the probability of detection of moving objects in Wide Area Motion Imagery (WAMI) that incorporates the effects of the target, the platform, and the environment. Developing situation awareness is vital for almost any kind of military operation. Through understanding the state and nature of the environment, military personnel can plan and respond accordingly. Situation awareness is often treated as the problem of knowing where all the potential targets are. Through knowing the locations of these targets, threats can be identified and countered. Another important source of awareness is to understand where targets cannot be. Regions that are free of targets can be used to constrain where targets might be. To meet these needs, Wide Area Surveillance (WAS) systems have been developed that are able to sense large swaths of an environment simultaneously and at high resolution. However, the next key challenge is to automatically analyze this image data to, for example, track the locations of targets and identify potential anomalous behavior. This report begins to explore how the output from a WAS system can be used by a state-of-the-art multi-target tracker. In particular, we considered how the output of the image processing and matching algorithms used in the Likelihood of Features Tracker (LoFT) could be combined with a Probabilistic Hypothesis Density (PHD) Filter. Using machine learning techniques, we have developed a formalism and algorithms to automatically predict how the visual appearance of a vehicle can change over time. Using this prediction model, we are then able to automatically threshold and detect potential candidate vehicle locations, and assess both probability of detection and the probability of clutter.