Event-RGB 융합을 기반으로 한 강건한 장소 인식 방법

Abstract

Event-based algorithms have been shown to have remarkable performance in computer vision. From the characteristics of event cameras that could provide a high dynamic range, event-based vision systems show robustness over appearance changes. Therefore, approaches based on event-based place recognition have been suggested. However, event sensors are not widespread, so event databases are not available in most target environments. This paper proposes to work around this problem by fusion with RGB measurements. From the point that the activation function of events is based on log magnitudes of intensity gradient, we propose to match event-generated edges to the edges generated from RGB images. Through experiments from public datasets, we show that event edges reconstructed from EventVLAD are matched into RGB images so that the image databases are compatible with event-based place recognition.