| With the rapid development of the aircraft, imaging techniques and computer vision, low-altitude aerial technology is widely used in aerial surveillance, disaster investigation and other military and civilian purposes. Taking into account the factor that the motion of imaging platform leading to the motion of stationary background and variability and complexity of shooting area, The common method of moving object detection(MOD) cannot achieve a good performance in the low-altitude aerial video, especially when there exist tall building, trees and mountains, etc., the height of these objects cannot be ignored compared with the distance of imaging platform. The big change angle between frames leads that these objects are projected onto image plane with a strong parallax, increasing interference to moving objects detection because the edge will be detected objects. Given that, we raise a fast and effective method of MOD in the presence of strong parallax in low-altitude aerial video. The main contribution of this paper is as follow:First, based on a deep analysis and research in low-altitude aerial video in the presence of strong parallax, we proposed a MOD framework in the video and give a brief introduction to key technologies in each step of the framework. The first step is image registration with the purpose of compensating the motion of the stationary background brought by the moving image platform. The second step is the initial moving object detection, roughly extracting moving object. The last step is parallax eliminating, eliminating the parallax caused by the existed tall object, detecting accurate moving object and extracting it completely.Second, with an in-depth analysis of the influence caused by moving imaging platform, variability and complexity of background and strong parallax, we propose an advanced image registration based on RANSAC, which can remove wrong feature matching features. Then we expound basic concepts of image registration and classification, give a detail introduction to three image feature-based registration method, which are Harris based image registration, SIFT based image registration and SURF based image registration. On the basis of three algorithms simulated in Matlab, we give a deep analysis and comparison of their performance to find the most suitable method for our paper. Last, given that the influence caused by strong parallax in coarse matching, leading to more wrong matching features, we proposed an epipolar constraint method based on advanced RANSAC to remove wrong feature matching features, which can improve the precision in feature matching and accuracy of image registration.Last, given that interference to MOD in low-altitude aerial video caused by strong parallax, we propose a fast and effective method based on gradient suppression and epipolar constrain. After image registration and initial MOD between two input frames, we apply gradient suppression remove parallax pixels first, and then use epipolar constrain to remove excess parallax pixels, which cannot be removed by gradient suppression. After applying these two method to remove parallax pixels, the result image will only exist little parallax pixels and most object pixels. We apply morphological operations to filtering out these pixels. And projection method will be employed to marking and classifying the moving object with the purpose of extracting the boundary and segmenting the whole moving object. |
