Robust Multi-View Target Attitude Determination Using Models, Multi-Cue Fusion and Machine Learning

Determination of maneuvering target attitude from ground-based imagery is complicated by the marginal resolutions, varying lighting conditions, cluttered backgrounds and poor target contrast. Fortunately, for cooperative tests, the external geometry of the missile can be known a priori and the location of the target missile can generally be localized in the image a priori as well. Based on these, Intelligent Fusion Technology, Inc., proposes a solution that takes the advantage of state-of-the-art segmentation algorithms, known 3D models of the target, and multiple views of the target in motion to maximize the overlay of the projections of target model on the image planes of viewing cameras and the measured scene images to robustly determine the attitude of the target. We put emphasis on the development of a benchmark data set as well as several benchmark methods for comparison against other state of the art approaches and methods. In particular, classical and deep learning-based segmentation algorithms with shape priors are proposed to battle with the highly cluttered backgrounds. Due to the lack of sufficient real data with flying targets, professional 3D modeling and animation tool is proposed to generate required data under varying conditions to facilitate the development of the proposed work.