Data-driven visibility enhancement using multi-camera system

Abstract

In bad weather conditions, with the presence of haze, fog or smoke, atmospheric particles attenuate the direct irradiance from the scene and scatter light to form airlight. Thus, visibility is decreased and may endanger important applications, such as outdoor surveillance or visual navigation for landing and taking off aircrafts. This paper proposes a novel method for visibility enhancement in bad weather conditions based on multi-view camera system. The main advantage of this method lies in the ability to solve ambiguities caused by texture-less, lack of color and contrast, while where most existing methods fail. The proposed system consists of two main components. First is a data-driven approach to extract template priors that are matched with current capturing dynamic scene images. A fixed multi-camera system is utilized to record dynamic scene appearances under different illuminations, in different time, seasons and weather conditions to construct the database which is explored to extract template models containing only static background objects and obtain corresponding scene structures in a data-driven manner. Second is dehazing based on current dynamic scene depth updated by fusing template depth with real-time multi-view stereo matching depth in foreground object regions. The proposed system achieves real-time and robust performances through combinations of data-driven prior extraction and dynamic scene depth optimization. Moreover, estimated weather condition parameters and the real-time reconstructed dynamic scene model are both useful byproducts. We believe that the proposed system is the first to dehaze based on multi-view camera system. An application based on airport surveillance demonstrates its effectiveness.