Multimodal classification of dance movements using body joint trajectories and step sounds

We present a multimodal approach to recognize isolated complex human body movements, namely Salsa dance steps. Our system exploits motion features extracted from 3D sub-trajec-tories of dancers' body-joints (deduced from Kinect depth-map sequences) using principal component analysis (PCA). These sub-trajectories are obtained thanks to a footstep impact detection module (from recordings of piezoelectric sensors installed on the dance floor). Two alternative classifiers are tested with the resulting PCA features, namely Gaussian mixture models and hidden Markov models (HMM). Our experiments on a multimodal Salsa dataset show that our approach is superior to a more traditionnal method. Using HMM classifiers with three hidden states, our system achieves a classification performance of 74% in F-measure when recognizing gestures among six possible classes, which outperforms the reference method by 11 percentage points.