LSTM-based real-time action detection and prediction in human motion streams

Motion capture data digitally represent human movements by sequences of 3D skeleton configurations. Such spatio-temporal data, often recorded in the stream-based nature, need to be efficiently processed to detect high-interest actions, for example, in human-computer interaction to understand hand gestures in real time. Alternatively, automatically annotated parts of a continuous stream can be persistently stored to become searchable, and thus reusable for future retrieval or pattern mining. In this paper, we focus on multi-label detection of user-specified actions in unsegmented sequences as well as continuous streams. In particular, we utilize the current advances in recurrent neural networks and adopt a unidirectional LSTM model to effectively encode the skeleton frames within the hidden network states. The model learns what subsequences of encoded frames belong to the specified action classes within the training phase. The learned representations of classes are then employed within the annotation phase to infer the probability that an incoming skeleton frame belongs to a given action class. The computed probabilities are finally compared against a learned threshold to automatically determine the beginnings and endings of actions. To further enhance the annotation accuracy, we utilize a bidirectional LSTM model to estimate class probabilities by considering not only the past frames but also the future ones. We extensively evaluate both the models on the three use cases of real-time stream annotation, offline annotation of long sequences, and early action detection and prediction. The experiments demonstrate that our models outperform the state of the art in effectiveness and are at least one order of magnitude more efficient, being able to annotate 10 k frames per second.