Automated action recognition is difficult because of the large number of possible meanings of data in video streams.
Many current approaches,train systems to look for specific actions. Those approaches are challenged when previously unseen actions are recognized, and often return inaccurate results if there are between actions from different classes in the training set.