Raindrop detection algorithm for ADAS

Abstract

Vision based advance driver assistance system (ADAS) completely rely on the images captured by the vehicle bound camera. Vision based ADAS also referred as computer vision based ADAS depends on clear images from surroundings for its algorithm implementation. If the images are hampered from external weather conditions such as rain or fog, the ADAS functionalities are affected. In general, checking the image, whether it is free from raindrop occlusions is one of the steps in the pre-processing, as clear, noise free images are required for ADAS algorithm deployments, which would enhance the ADAS functionalities. In ADAS, generically raindrop occlusions is identified by comparing two successive images then sorting analysis using displacement formulae or by identifying them typically based on photometric or geometric properties of raindrop. The computation time and the accuracy for these approaches are trade off factors. In this paper, we propose an algorithm to detect the raindrop, where detection is performed using thresholding and feature transform only on single frame. The approach reduces the computation time and memory resource, executing only on current frame. Once the raindrop occlusion is affirmed, an interrupt is sent to the concurrent algorithm about the occlusions and abort the same. To retrieve the region from occlusion we follow one of the generic method employing Gauss filter.