단일카메라를 이용한 동적 상황에서의 추돌방지를 위한 움직이는 물체 검출 시스템

Abstract

We proposed an background model using feature point's position ratio in image plane for moving object detection. The proposed model require monocular camera and non-tracking based assumption. The moving object detection algorithm using the proposed model can detect various dangerous moving object which approaching to the camera or crossing the camera. When camera moves straight forward, the proposed model assume that the background feature point can estimated by previous frames and lies on the epipolar line. To satisfy straight forward condition, rotation compensation are needed. We used SIFT algorithm for feature point detection and 5-parameter model for rotation compensation. In experiments on challenging KITTI dataset, we compared our algorithm with scene flow algorithm that uses stereo camera and achieved meaningful detection ratio.