Parallax removing for ground moving target detection on non-flat terrain

Ground moving target detection is one of the major challenges in the field of aerial video surveillance. Most of the existing methods introduce a large number of parallax false alarms while detecting moving objects. In this case, we propose a novel method, which fuses clues from 3D scene structure and low level image motion for joint detection. To achieve parallax removing capability with satisfying performance, the proposed method eliminates the discontinuities parallax by flow association. This step remove a significant portion of the parallax, but there are still false alarms which are close to the ground motion targets. In the last step, epipolar constraint is conducted to final motion analysis, which stands for the information from 3D scene. To verify the feasibility and effectiveness of our algorithm, we carry out large amounts of experiments on aerial videos captured in complex urban scene. Extensive experimental results show that our detection algorithm can remarkably reduce parallax false alarms.