Exemplar-based Human Interaction Recognition: Features and Key Pose Sequence Model

Due to intra-class variation, camera jitter, background clutter, etc, human activity recognition is a challenging task in computer vision. We propose an exemplar-based key pose sequence model for human interaction recognition. In our model, an activity is modelled with a sequence of key poses, important atomic-level actions performed by the actors. We employ a strict temporal ordering of the key poses for each actor, an exemplar representation is used to model the variability in the instantiation of key poses. To utilize interaction information, spatial arrangements between the actors are included in the model. Quantitative results that form a new state-of-the-art on the benchmark UT-Interaction dataset are presented. Results on a subset of the TRECVID dataset are also promising.