Vehicle Tracking in UAV-Captured Video Data Incorporating Adaptive Weighting of Features

Abstract : In our previous work we have investigated data fusion frameworks for object detection and tracking. In particular, we proposed the use of the spatiogram, as opposed to the more commonly used histogram, for object tracking. The motivation for this stems from the fact that a spatiogram allows coarse encoding of an object's spatial information. In this way it is more robust to changes in object appearance than a histogram that carries no spatial information, whilst not suffering the constraints of template-based matching techniques that encode full object spatial information. In addition, we proposed an improved similarity measure for matching spatiograms from one frame to the next. Furthermore, we embedded this measure in an innovative efficient mean-shift search process. Using this as a basis, we were able to propose a multi-feature tracking framework that we term a spatiogram bank that leverages earlier work on optimal feature fusion. We proved that the spatiogram is particularly suited to this framework as it does not suffer from the curse of dimensionality as new features are added, leading to a very flexible multi-feature tracking framework. We demonstrated the benefits of this framework in a variety of different application scenarios that use visible spectrum and infrared visual data sources. In particular, we showed how this tracking framework can be embedded in a system for detection and tracking of ground-based vehicles in UAV video footage.