This work presents a spatio-temporal motion descriptor that is computed from a spatially-constrained decomposition and applied to online classification and recognition of human activities. The method starts by computing a multi-scale dense optical flow that provides instantaneous velocity information for every pixel without explicit spatial regularization. Potential human actions are detected at each frame as spatially consistent moving regions and marked as Regions of Interest (RoIs). Each of these RoIs is then sequentially partitioned to obtain a spatial representation of small overlapped subregions with different sizes. Each of these region parts is characterized by a set of flow orientation histograms. A particular RoI is then described along the time by a set of recursively calculated statistics, that collect information from the temporal history of orientation histograms, to form the action descriptor. At any time, the whole descriptor can be extracted and labelled by a previously trained support vector machine. The method was evaluated using three different public datasets: (1) The VISOR dataset was used for two purposes: first, for global classification of short sequences containing individual actions, a task for which the method reached an average accuracy of 95% (sequence rate). Also, this dataset was used for recognition of multiple actions in long sequences, achieving an average per-frame accuracy of 92.3%. (2) the KTH dataset was used for global classification of activities and (3) the UT-datasets were used for evaluating the recognition task, obtaining an average accuracy of 80% (frame rate).