INRIA-WILLOW at TRECVid 2010: Surveillance Event Detection

This paper presents a system evaluated in the Surveillance Event De- tection (SED) task of TRECVid 2010 campaign. We investigate a generic sta- tistical approach applied to seven event classes defined by the SED task. Our video representation is based on local space-time descriptors which are vector- quantized and aggregated into histograms within short temporal windows and spatial regions defined by the prior. We use priors on the spatial localization of actions estimated from the spatio-temporal annotation of actions in the training data. To recognize actions, we learn one-against-all action classifiers using non- linear SVMs. Each classifier is applied independently to localize temporal inter- vals of actions using window-scanning approach. We present results of six runs with variations in the two parameters: (i) classifier threshold and (ii) temporal extent of the scanning window.