Fast realistic multi-action recognition using mined dense spatio-temporal features