The use of sparse plus low-rank decomposition on moving object and change detection in videos

Abstract

A solution for the detecion of anomalies, such as change and moving objects in videos, is to obtain a low-rank representation of the frames that compose the video sequence and try to reconstruct a frame from the original video using a combination of the low-rank representation of the others. This thesis propose algorithms that project the low-rank structures into a low-dimensional union-of-subspaces, to solve this problem allowing the model to cope with dynamic backgrounds such as those found in videos acquired with moving cameras and other complex scenarios. Part of the thesis covers change detection in videos acquired with moving cameras. The proposed algorithms provide good detection results, at the same time as obviate the need for previous video synchronization. They also use properties of the data representation in order to restrict the search space to the most relevant subspaces, providing computational complexity gains of up to 100 times and 91% true positive and only 33% false positive detections on experiments using the VDAO database, with abandoned objects in a cluttered industrial scenario. Another part presents a solution to the detection of moving objects in the presence of highly dynamic backgrounds. The proposed solutions use low-rank and sparse matrix decompositions to represent the background as a union-of-subspaces, while applying saliency maps to restrict the updates of the foreground matrix. The proposed methods presents low false positive detection rate, and is shown to achieve state-of-the-art performance among similar methods, attaining 0.74 F1 score in the UCSD dataset.