Activity recognition using logical hidden semi-Markov models
Yabing Zha, Shiguang Yue, Quanjun Yin, Xiaocheng Liu · 2013
Activity recognition is challenging and valuable in both real and virtual world. As important directed graphical models, hidden Markov models and their extensions are widely used to solve probabilistic activity recognition problems. In this paper, logical hidden semi-Markov models (LHSMMs) which combine logical hidden Markov models (LHMMs), a statistical relational learning method, and hidden semi-Markov models are proposed, and the lognormal distribution is used to model the duration explicitly. The formal description of LHSMMs and the exact inference process using a logical forward algorithm with duration are presented; the directed graphical representation of unmanned aerial vehicle activities is also given. Experiments are also designed to compare the performances of LHSMMs and LHMMs. The results prove that, the recognition result of abstract states using LHSMMs is more smoothing, and the probability of the real instantiated activity is larger than that of LHMMs in most time because of modeling duration explicitly.