Towards Intelligent Mission Profiles of Micro Air Vehicles: Multiscale Viterbi Classification

In this paper, we present a vision system for object recognition in aerial images, which enables broader mission profiles for Micro Air Vehicles (MAVs). The most important factors that inform our design choices are: real-time constraints, robustness to video noise, and complexity of object appearances. As such, we first propose the HSI color space and the Complex Wavelet Transform (CWT) as a set of sufficiently discriminating features. For each feature, we then build tree-structured belief networks (TSBNs) as our underlying statistical models of object appearances. To perform object recognition, we develop the novel multiscale Viterbi classification (MSVC) algorithm, as an improvement to multiscale Bayesian classification (MSBC). Next, we show how to globally optimize MSVC with respect to the feature set, using an adaptive feature selection algorithm. Finally, we discuss context-based object recognition, where visual contexts help to disambiguate the identity of an object despite the relative poverty of scene detail in flight images, and obviate the need for an exhaustive search of objects over various scales and locations in the image. Experimental results show that the proposed system achieves smaller classification error and fewer false positives than systems using the MSBC paradigm on challenging real-world test images.