iCAST: Impact of Climate on Assistive Scene Text detection for autonomous vehicles

Scene-text detection uses machine learning and computer vision methods in order to automate the process of text detection in natural environmental conditions. The study of scene text detection becomes more critical in the context of self-driving cars, where text present on traffic panels, hoardings, and other boards and signs needs to be detected and recognized from the camera installed in the car. The variations in the car's surrounding environment, where text is in different orientations, languages, angles, and lighting and weather conditions, makes this task more complex. This paper investigates the impact of different weather conditions by applying haze removal techniques on RoadText-1k dataset in preprocessing. These techniques have improved qualitative and quantitative results in terms of recall metric.