Automatic extraction of moving objects from UAV-borne monocular images using multi-view geometric constraints

This paper proposes a method to detect dynamic objects in the images obtained by a small UAV. Two geometric constraints in multi-view images are used to classify each of the extracted featurepoints as static or dynamic. The first constraint is the epipolar constraint which requires static points to lie on the corresponding epipolar lines in the subsequent image. The second constraint, named as flow-vector bound constraint here, restricts the motion of static points along the epipolar lines. In addition, the pose of the UAV-borne camera, which is required when applying these constraints, is estimated by using a vision-based SLAM method, PTAM. The proposed method fully exploits the characteristics of UAV-borne images and achieves satisfactory results. The algorithms were tested with a small quadrotor platform in a real-world scene and successfully detected features extracted from multiple pedestrians.