Geometric distortion correction of long-range imaging containing moving objects

Abstract

Video sequences of long-range imaging are inevitably affected by optical turbulence, which leads to non-uniform geometric distortion and object position shifts. The restoration of this turbulence-degraded data containing moving objects is a challenging task, which essentially involves video stabilization and the detection of moving objects. In this work, a novel method is proposed for simultaneously realizing turbulence mitigation and moving objects detection. We firstly model the turbulent foreground with a specific mixture of Gaussian (MoG) distribution, which is regularized online by the low-rank subspace of the background. Then, to preserve the low-rank property of the dynamic background, we embed a transformation operator into the proposed model which makes it much more robust in the event of practical camera jitters or rotation. Finally, a simple mask strategy is used to reconstruct stable frames containing moving objects. Extensive experiments using synthetic and real-life turbulence-degraded data show that the proposed method outperforms other compared approaches in terms of both geometric distortion correction and the preservation of moving objects.