Dynamic Foreground/Background extraction based on segmented image

10 This paper addresses the problem of extracting dynamic foreground regions 11 from a relatively complex environment within a collection of images or a 12 video sequence. By using image segmentation code, we can first convert our 13 traditional pixel-wise image collection into a collection of image with 14 multiple monochrome image segments. Our approach in this study consists of 15 four steps. First of all, we uniformly extract patches from the first frame of 16 segmented image collection. And then, we manual tag the foreground and 17 background within the first segmented image. After this step all the patches in 18 the first frame will form two bags of patches. For one bag, the patches in it 19 can model the features of the foreground. Meanwhile, the patches in the other 20 bag can describe the features of the background. In this case, we call them 21 foreground patches and background patches respectively. Third, for an 22 incoming frame, we perform the segmentation and then extract the patches. 23 For both the patches from the new frame and the previous known patches, in 24 order to reduce the dimension, we perform LDA. Then use KNN and KDE to 25 find the nearest patch bag for the incoming patches. At this point, we can 26 differentiate the foreground and background for the new image frame. Finally, 27 we perform a bidirectional consistency check between the patches we already 28 get and the patches from incoming image frame make the model adapt to new 29 incoming image frames. In this report, we practice a novel, clear and easy 30 way to extract dynamic foreground from the complex background. 31 32