Skip to Main Content
In this paper, we describe a novel technique for detecting raindrops using in-vehicle camera images. The appearance of raindrops on a car windshield can depend on their background, so it is often difficult to detect them using conventional template matching methods, which are based on image features. Initially, we extract potential raindrop regions from images, before generating a rendered background image using a physical raindrop model based on the refraction of light rays. This rendered image is then used to identify true raindrops based on their similarity to the true background image. We propose a new model that approximates a raindrop shape as a spheroid section. This method can represent different raindrop shapes more adaptively and flexibly than conventional models, which approximate raindrops as a section of a sphere. We also extend the Maximally Stable External Regions algorithm to extract candidate raindrops and we identify three measures of image similarity using a Support Vector Machine algorithm. We conducted experiments that confirmed the effectiveness of the proposed technique.