Detection of Camera Artifacts from Camera Images

Cameras are frequently used in state-of-the-art systems in order to get detailed information of the environment. However, when cameras are used outdoors they easily get dirty or scratches on the lens which leads to image artifacts that can deteriorate a system's performance. Little work has yet been done on how to detect and cope with such artifacts. Most of the previous work has concentrated on detecting a specific artifact like rain drops on the windshield of a car. In this paper, we show that on moving systems most artifacts can be detected by analyzing the frames in a stream of images from a camera for static image parts. Based on the observation that most artifacts are temporally stable in their position in the image we compute pixel-wise correlations between images. Since the system is moving the static artifacts will lead to a high correlation value while the pixels showing only scene elemets will have a low correlation value. For testing this novel algorithm, we recorded some outdoor data with the three different artifacts: raindrops, dirt and scratches. The results of our new algorithm on this data show, that it reliably detects all three artifacts. Moreover, we show that our algorithm can be implemented efficiently by means of box-filters which allows it to be used as a self-checking routine running in background on low-power systems such as autonomous field robots or advanced driver assistant systems on a vehicle.