A novel method for quick on-line segmentation based on sparsity

We propose a novel multi-estimate dynamic programming (DP) method for on-line detection and segmentation of significant anomalies in a video sequence. The method is based on the concept of sparsity, which means that we reduce visual features of each frame to a set of keypoints. In our linescan application this is done by extracting only the intensity extrema. This way, we can decrease DP´s inherent noise-amplifying tendency when building up the estimates of an anomaly. For detection improving, we introduce weights that express similarity between the spatial distribution of pixels forming so-called DP score sums and a reference representing their assumed distribution. The spatial dynamics estimation is improved by 30 % if compared to the intensity-only DP. Some 59 % change point recovery rate is attained in a web imaging application where illumination varies, contrasts are small, and the decision making time is limited to fractions of a second due to high-speed running of the web.