Moving object detection under moving camera by using normal flows

Ding Yuan, Yalong Yu

Abstract: Moving object detection under a moving camera has long been a very challenging task as the movements of the moving objects and motion caused by the camera are merged in the image sequence. Traditional methods tend to tackle this problem by establishing the matching features across the image frames, or by inferring optical flows. However, both matching features and calculating optical flows are ill-posed problems in the computer vision. Some artificial assumptions, for example smoothness assumption and continuity assumption, have to be made to get the solutions. Such assumptions describing an extremely ideal capturing scene usually do not hold. In this paper, we present a novel algorithm on detecting the moving object under a moving camera by using normal flows which can be calculated directly from the image sequence data without any artificial constraints. Therefore our novel method is completely independent of the structure of the captured scene or object. We tested our algorithm on both synthetic image data and real image sequences, and the experimental results showed the good performance of our algorithm.