Fast Video Saliency Detection Based On Spatiotemporal Cues

In recent years, image saliency detection technology is developing rapidly and has attracted a lot of researchers. While video saliency detection technology, usually regarded as the extension of image saliency detection in computer vision field, just emerges in several years and is still in the early stages. As video saliency detection is incorporated with time information, it is considered as a very challenging subject.The article proposed an algorithm which processes input videos from both space and time cues. In terms of space, the algorithm performs static saliency detection via graph-based manifold ranking aiming at the low layer information including color and intensity etc. and performs fast object detection based on Haar-like features aiming at high layer information like human face. In terms of time, the algorithm performs fast dynamic saliency detection based on fast and dense optical flow analysis aiming at the motion information in the frame sequence. In addition, the article proposed a reasonable and effective fusion mechanism to integrate the result computed by the above cues, and get the final saliency map of each frame. The article verifies the effectiveness of the algorithm from four different aspects in the experimental part, and presents two different kinds of objectively and reasonable evaluation algorithm for video saliency detection. At last the algorithm is realized in the Omron dataset and performs good result. Some idea which helps to improve the algorithm is also presented, aiming at some existing problems.Since the data size is usually very huge in a video, the algorithm will lose its significant of practical application if the processing speed is too slow. The article regards the fast processing speed as an important premise and gains good result, which can attain the5frames per second.