We classify human actions occurring in depth image sequences using features based on skeletal joint positions. The action classes are represented by a multi-level Hierarchical Dirichlet Process -Hidden Markov Model (HDP-HMM). The non-parametric HDP-HMM allows the inference of hidden states automatically from training data. The model parameters of each class are formulated as transformations from a shared base distribution, thus promoting the use of unlabelled examples during training and borrowing information across action classes. Further, the parameters are learnt in a discriminative way. We use a normalized gamma process representation of HDP and margin based likelihood functions for this purpose. We sample parameters from the complex posterior distribution induced by our discriminative likelihood function using elliptical slice sampling. Experiments with two different datasets show that action class models learnt using our technique produce good classification results.

The problem of classifying human activities occurring in depth image sequences is addressed. The 3D joint positions of a human skeleton and the local depth image pattern around these joint positions define the features. A two level hierarchical Hidden Markov Model (H-HMM), with independent Markov chains for the joint positions and depth image pattern, is used to model the features. The states corresponding to the H-HMM bottom level characterize the granular poses while the top level characterizes the coarser actions associated with the activities. Further, the H-HMM is based on a Hierarchical Dirichlet Process (HDP), and is fully non-parametric with the number of pose and action states inferred automatically from data. This is a significant advantage over classical HMM and its extensions. In order to perform classification, the relationships between the actions and the activity labels are captured using multinomial logistic regression. The proposed inference procedure ensures alignment of actions from activities with similar labels. Our construction enables information sharing, allows incorporation of unlabelled examples and provides a flexible factorized representation to include multiple data channels. Experiments with multiple real world datasets show the efficacy of our classification approach.

No abstract

No abstract

No abstract