In video surveillance, classication of visual data can be very hard, due to the scarce resolution and the noise characterizing the sensors' data. In this paper, we propose a novel feature, the ARray of COvariances (ARCO), and a multi-class classifcation framework operatingon Riemannian manifolds. ARCO is composed by a structure of covariance matrices of image features, able to extract information from data at prohibitive low resolutions. The proposed classication framework consists in instantiating a new multi-class boosting method, working on the manifold Sym+ of symmetric positive defnite d x d (covariance) matrices. As practical applications, we consider different surveillance tasks, such as head pose classifcation and pedestrian detection, providing novel state-of-the-art performances on standard datasets.
Tosato, D., Farenzena, M., Cristani, M., Spera, M., Murino, V., Multi-class Classification on Riemannian manifolds for Video Surveillance, Computer Vision - ECCV 2010, Springer, Berlino 2010 <<LECTURE NOTES IN COMPUTER SCIENCE>>, 6312: 378-391 [http://hdl.handle.net/10807/35679]
Multi-class Classification on Riemannian manifolds for Video Surveillance
Spera, Mauro;
2010
Abstract
In video surveillance, classication of visual data can be very hard, due to the scarce resolution and the noise characterizing the sensors' data. In this paper, we propose a novel feature, the ARray of COvariances (ARCO), and a multi-class classifcation framework operatingon Riemannian manifolds. ARCO is composed by a structure of covariance matrices of image features, able to extract information from data at prohibitive low resolutions. The proposed classication framework consists in instantiating a new multi-class boosting method, working on the manifold Sym+ of symmetric positive defnite d x d (covariance) matrices. As practical applications, we consider different surveillance tasks, such as head pose classifcation and pedestrian detection, providing novel state-of-the-art performances on standard datasets.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.