The analysis of haze effect on dense semantic mapping

This paper addresses the issue of dense semantic mapping in hazy scenes. In the past few decades, extensive research has been performed on semantic mapping in clear scenes. However, there was little attention on dense semantic mapping in hazy environments. In this paper, we try to solve this problem. Towards this aim, we introduce a hazy dataset which is built on the TUM dataset. In order to explore the haze effect on dense semantic mapping, we have performed a lot of experiments and evaluated several state-of-the-art dehazing methods. In addition, we adopt a convolutional neural network (CNN) for image preprocessing to improve the robustness of robot localization and mapping in hazy scenes. The experimental results show that a good dehazing method can effectively reduce the tracking failure of simultaneous localization and mapping (SLAM) in hazy scenes and benefit semantic understanding.