People silhouette extraction from people detection bounding boxes in images

Abstract

In many applications such as video surveillance or autonomous vehicles, people detection is a key element, often based on feature extraction and combined with supervised classification. Usually, output of these methods is in the form of a bounding-box containing an extracted people along with the background. But in specific application contexts, this bounding box information is not sufficient and a precise segmentation of people silhouette is needed inside the bounding box. For videos, this is actually solved by using background subtraction strategies. However, this cannot be considered for the case of still images that also occur in many video surveillance applications. To that aim, we propose to consider that issue in this paper. The principle is to devise a complete scheme for people segmentation inside people detection bounding boxes. Such a scheme relies on several steps: pre-processing, feature extraction and probability map computation to approximately locate people silhouette, and graph cut clustering to refine the silhouette from the map prior. Since many different methods can be considered, along with their associated parameters, tuning, we use a systematic approach towards determining the best combination scheme to conceive a segmentation scheme. The F-measure is used as a benchmark for evaluation. Experimental results show the benefit of the proposed approach that goes beyond the actual state-of-the-art.