In this paper we presented a violence detector built on the concept of visual codebooks using linear support vector machines. It differs from the existing works of violence detection in what concern the data representation, as none has considered local spatio-temporal features with bags of visual words. An evaluation of the importance of local spatio-temporal features for characterizing the multimedia content is conducted through the cross-validation method. The results obtained confirm that motion patterns are crucial to distinguish violence from regular activities in comparison with visual descriptors that rely solely on the space domain.