Abstract
Video desnowing/deraining plays a vital role in outdoor vision systems, such as autonomous driving and surveillance systems, since the weather conditions significantly degrade their performance. Although numerous approaches have been reported for video snow/rain removal, they are limited to a few videos and did not consider the variations that occurred for the camera and background in real applications. We build a complete snow and rain dataset to overcome this limitation, consisting of 577 videos with synthetic snow and rain, quasi-snow, and real snow and rain. All possible variations of the background and the camera are considered in the dataset. Then, an efficient pixel-wise video desnowing/deraining method is proposed based on the color and temporal information in consecutive video frames. It is highly likely for a single pixel to be a background pixel rather than a snowy pixel at least once in the consecutive frames. Inspiring from this fact along with the color information of the snow pixels, we extract the background pixels from different consecutive frames by searching for the minimum gray-scale intensity. Experimental results demonstrate and validate the proposed method’s robustness to illumination and high-performance static background and camera.