Machine learning method based on a bag of key poses model and temporal alignment.
This method handles multiclass classification of data with sequential or temporal relation. Any type of feature expressed as an array of double values can be used. This data has to be acquired in an ordered fashion, so that a sequence of features with a meaningful order can be obtained.
This method has successfully been employed for real-time recognition of human actions. In this case, spatial information is provided in the form of frame-based features that describe the human pose, whereas sequences of these features encode the temporal evolution. Therefore, this learning method applies spatio-temporal classification by means of learning the bag of key poses model and applying temporal alignment between the test sequence and the previously learned templates.