Online action recognition is nowadays a major challenge on computer vision due to uncontrolled scenarios, variability on dynamic action representations, unrestricted capture protocols among many other variations. This work introduces a very compact binary occurrence motion descriptor that allows to recognize actions on partial video-sequences. The proposed approach starts by computing a set of motion trajectories that represent the developed activity. On that regard, a local counting process is performed over bounded regions, and centered at each trajectory, to search for a minimal number of neighboring trajectories. This process is then codified in a vector of binary values (ToBPs) that will create a regional description, at any time of the video sequence, to represent actions. This regional description is obtained by determining the most recurrent binary descriptors in a particular video interval. The final regional descriptor is mapped to a machine learning algorithm to obtain a recognition. The proposed strategy was evaluated on three public datasets, achieving an average accuracy of 70% in tasks of action recognition by using a local descriptor of only 51 values and a regional descriptor of 400 values. This compact description constitute an ideal condition for real-time video applications. The proposed approach achieves a partial recognition above 70% on average accuracy using only the 40% of videos.
CITATION STYLE
Garzón, G., & Martínez, F. (2020). Online Action Recognition from Trajectory Occurrence Binary Patterns (ToBPs). In Advances in Intelligent Systems and Computing (Vol. 1066, pp. 409–418). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-030-32022-5_38
Mendeley helps you to discover research relevant for your work.