ABSTRACT
We present a system to detect and track moving objects from an airborne platform. This application involves many important techniques in computer vision. We fi rst discuss the motion detection on a moving platform. A sliding window-based method is used to build the dynamic background model of an airborne camera. This approach is robust to accumulated errors in image registration. We then present a 2D geo-registration method to register UVA images with a satellite image, where motion model of objects is more physically meaningful than in image coordinates. Mutual information is used to fi nd correspondences between these two different modalities. After motion detection and geo-registration, tracking is performed in a hierarchical manner: at the temporally local level, moving image blobs acquired by motion segmentation are associated into tracklets by using the Markoc chain Monte Carlo data association (MCMC-DA) algorithm; at the global level, tracklets are linked by their appearance and spatiotemporal consistence on the global map. To achieve effi cient time performance, graphics processing unit (GPU) techniques are applied in motion detection, which is the bottleneck of the whole system.
