This work introduces a novel motion descriptor that enables human activity classification in video-surveillance applications. The method starts by computing a dense optical flow, providing instantaneous velocity information for every pixel. The obtained flow is then characterized by a per-frame orientation histogram, weighted by the norm, with orientations quantized to 32 principal directions. Finally, a set of global characteristics is determined from the temporal series obtained from each histogram bin, forming a descriptor vector. The method was evaluated using a 192-dimensional descriptor with the classical Weizmann action dataset, obtaining an average accuracy of 95 %. For more complex surveillance scenarios, the method was assessed with the VISOR dataset, achieving a 96.7 % of accuracy in a classification task performed using a Support Vector Machine (SVM) classifier.