An Efficient Combination Between Epipolar Geometry and Perspective Transform for Moving Object Detection

An efficient moving object detection (MOD) method for monocular camera with ego-motion is proposed in this paper. In general, there are keypoint-based foreground detection methods that find out outliers using epipolar geometry of two camera coordinate systems in consecutive frames. Determining of regions of moving objects from sparse keypoints is especially challenging. In this paper, we propose an optical flow noise removal and an efficient sparse-to-dense motion probability map. Subtraction method using perspective transform is applied to detect candidate regions of objects. Moving objects are detected by combining subtraction method with the proposed motion probability map. For stable detection, we use the Kanade-LucasTomasi (KLT) feature tracker for keypoint tracking and Kernelized Correlation Filter (KCF) tracking algorithms for bounding boxes. Through the simulations, we show that our proposed probability map is effective to detect various sizes of objects.