HDP-HMM-SCFG: A Novel Model for Trajectory Representation and Classification

Weiguang Xu, Yafei Zhang, Jianjiang Lu, Jiabao Wang
2011 Procedia Engineering  
In this paper, we propose a novel model, HDP-HMM-SCFG, for representing and classifying trajectories. Trajectories are represented by stochastic grammar, where trajectory segments are considered as observations emitted by the grammar terminals, which are attached with HMMs. In order to learn the parameters of SCFG, we employ hierarchical Dirichlet Processes (HDP) as the nonparameter prior of the distribution of the parameters, and obtain the model of HDP-HMM-SCFG. Then, we propose a 3-level CRP
more » ... based Gibbs sampling inference algorithm to acquire the SCFG parameters. In the training phase of classification, SCFGs for different classes are learned respectively by inferring on the training sets with different labels independently. Then test trajectory is parsed with a bottom-up parsing algorithm, and the probability for each SCFG to generate it is calculated. The label of the class with the maximum likelihood to generate the test trajectory is assigned to the test trajectory. Experiment on ASL dataset is carried on to validate our approach.
doi:10.1016/j.proeng.2011.08.117 fatcat:e3welr5lurebpainfdkwjrv7di