Analysis of Multi-planar Probability Maps for People Localization in Overlapping Camera Systems
Résumé
This paper proposes a novel method for people localization in multi-camera video surveillance systems. Multi-planar projections of overlapping views are fused to obtain probability maps known as synergy maps. However, synergy maps suffer from false detections known as ghosts, and the purpose of this article is to use pattern recognition techniques for ghost pruning. To this end, the geometry of the system is used to generate synthesized synergy maps corresponding to the presence of a person at precise locations. These ideal patterns are then compared to the real synergy map using a discriminative measure in order to decide whether a person is present or not. This approach is novel because rather than a priori decision rules to remove the ghosts, we propose to compare the physical reality of the scene with the synthetic data, and thus, validate or not the assumptions of people detection using pattern matching methods. The proposed method achieves excellent results on the PETS 2009 dataset as demonstrated by comparison with state-of-the-art methods.