Switching GMM-HMM for Complex Human Activity Modeling and Recognition

Complex human activities can be decomposed into primitive activities (PAs) that happen sequentially but may vary in order or frequency among different observation sequences. The Gaussian mixture model based hidden Markov model (GMM-HMM) is widely used for activity modeling but is only capable of modeling activities with fixed pattern trajectories. To overcome the drawback of GMM-HMM, a hierarchical structure is introduced to form the Switching Gaussian mixture model based hidden Markov model (S-GMMHMM), where PAs are treated as upper layer states and the continuous observation sequence emitted by each PA is modeled by a GMM-HMM. To ensure that the upper layer states correspond to real activities, a supervised algorithm is proposed for S-GMMHMM parameter estimation. For the purpose of less time complexity, a real-time activity recognition algorithm is proposed by computing activity posteriors recursively. Experiment results show that the proposed model outperforms GMM-HMM in activity recognition, while brings a notable reduction in time complexity.