People-background segmentation with unequal error cost

Abstract

Esta es la versión de autor de la comunicación de congreso publicada en: This is an author produced version of a paper published in: ABSTRACTWe address the problem of segmenting a video in two classes of different semantic value, namely background and people, with the goal of guaranteeing that no people (or body parts) are classified as background. Body parts classified as background are given a higher classification error cost (segmentation with bias on background), as opposed to traditional approaches focused on people detection. To generate the people-background segmentation mask, the proposed approach first combines detection confidence maps of body parts and then extends them in order to derive a background mask, which is finally post-processed using morphological operators. Experiments validate the performance of our algorithm in different complex indoor and outdoor scenes with both static and moving cameras.

Post-processing approaches for improving people detection performance

“…The first one is based on the people-background segmentation [51]. The second one is the combination or fusion of up to six independent, appearance based people detectors at decision-level and their combination with a motion based people detector.…”

“…The corresponding background segmentation mask is finally generated after binarization and post-processing. Therefore, although the people-background segmentation [51] is based on [41], the objective and result are fairly different from the traditional people detection approaches. …”

“…People-background segmentation [51] is a two-class segmentation ensuring that no people or body parts are appearing in the background class. This type of segmentation is useful not only as a people detection preprocessing or post-processing step, but also for other video analysis processes such as tracking and people density estimation.…”

“…Firstly, people detections could be obtained using any people detector from the state of the art and the people-background segmentation is obtained with the approach proposed in [51] (see previous section 3.1). Then, both information sources are combined with the aim of eliminating or reducing the number of false detections while keeping, as much as possible, the number of positive detections.…”

“…The chosen people-background segmentation method [51] uses the DTDP detector [41] in order to detect different body parts and extends this representation by appropriately grouping them. Then, they fuse detection confidence maps according to regions that are expected to be covered by the body parts.…”