Evaluation of Background Subtraction Algorithms Using Muhavi, a Multicamera Human Action Video Dataset

MuHAVi is a human action video dataset that provides manually annotated silhouette data for evaluating action recognition methods that are based on silhouettes. The manual silhouette data is intended to separate segmentation difficulties from action recognition performance evaluation. However, the number of annotated images is limited to a small set of sequences and to generate new manually annotated silhouettes is a very complex and costly process. We propose to use state of art segmentation algorithms that automatically generate silhouettes for the whole dataset. This paper addresses the performance evaluation of algorithms that separate foreground and background in the MuHAVi dataset, particularly for methods that create a statistical model of background. It discusses background subtraction algorithms based in mixture of Gaussians, and presents the metrics commonly used in segmentation evaluation. The best evaluated algorithm is used to automatically generate a set of silhouettes. © The Institution of Engineering and Technology 2014.