This paper presents a robust, accurate and real-time model-based tracking method for markerless objects in complex environments to replace the conventional 3D tracking approach based on cooperative targets. A known 3D model of the object is projected onto a 2D plane and occlusion culling is performed with the precalibrated intrinsic parameters and initialized pose. The correspondences between a 3D object model and 2D image edges are commonly used to estimate the camera pose, so the pose optimization problem is transformed into 3D/2D model-to-image registration. For each visible model sample point, a one-dimensional search for putative image edge points is then performed along a direction perpendicular to its line by state-of-the-art methods. However, false correspondences always occur due to cluttered backgrounds or partial occlusion. To overcome this problem, a new search scheme for obtaining line correspondences instead of edge point correspondences is implemented. The outliers of 3D/2D line correspondences are then effectively detected and removed with algebraic outlier rejection, where the camera pose is iteratively optimized from correct correspondences of 3D/2D lines by minimizing the perpendicular distances from the endpoints of 3D model lines to their corresponding projection planes. The method presented in this paper has been validated on both synthetic images and real data. The experimental results show that the method is robust to strong noise, exquisite illumination changes and highly cluttered backgrounds. Meanwhile, it can easily satisfy the real time request.
Micro satellites have been widely used in the target monitoring and tracking. Aimed to reduce the ground operator's workload, a target location method based on homography and scene matching is proposed in this paper. For the first time satellite flies over target area, it needs the operator to take a frame as a reference image and extract the target area which is regarded as plane scene. When the satellite scan the area again, we take a frame as a real-time image and calculate the homography induced by the plane. Then rectify the reference image with the homography to reduce distortion between the two images. Finally, locate target on the real-time image using matching method. A significant-feature-point auxiliary positioning method is also proposed to adapt to target area without obvious features. It adopts affine model to calculated target location on the real-time image. Simulation experimental results show accuracy and practical value for engineering of the proposal method.