Adherent Raindrop Detection

Abstract: Raindrops adhered to the glass protecting a surveillance camera can significantly degrade the visibility of a scene. The goal of this master's thesis is to develop an accurate, efficient, computationally cheap algorithm that automatically detects adherent raindrops using only video and then removes them by activating the wipers or the shaking dome function. Two already existing algorithms, Temporal Intensity Difference and Maximally Stable Extremal Regions, were tested. Furthermore, additional criteria were added to the existing algorithms to improve performance, such as requiring the raindrops to be detected when they land on the screen. An algorithm that detects lens flares was also developed. The algorithms were tested on common surveillance scenes during both day and night. Combining these criteria and algorithms proved to be better than already existing methods and still fulfils the requirement that the algorithms should be computationally cheap. The final algorithm on average found 49% of the drops present in the picture with a false detection rate of 6.7% and took on average 3.1 seconds to run.