Pose based activity recognition using Multiple Kernel learning

We describe a method for activity recognition based on distribution of human poses in a video. Pose estimation has shown to be sensitive to the priors given to the inference method; we use a collection of distinctive kinematic tree priors to model the variety of pose variations present in a video. Feature histograms are computed from vector quantized descriptors derived from the pose estimates. A learned Multiple Kernel SVM classifier is used to combine the various histograms to give activity classifications. We report results on a publicly available human gesture dataset.