Place recognition based on Latent Dirichlet Allocation using Markov chain Monte Carlo method

Place or scene recognition is an important competence of mobile robots for operating in a real dynamic environment. Latent Dirichlet Allocation (LDA), a popular probabilistic model, can achieve outstanding performance in image recognition, and it has been attracted the attention of a large number of researchers. Parameter estimation is a key step for the learning procedure of LDA. In this paper, we propose a novel place recognition approach based on LDA using Markov chain Monte Carlo (MCMC) for approximate inference technique. Firstly, the training images of each category are represented as a set of different themes. And MCMC is employed to estimate parameters of LDA instead of variational inference. Then an unknown test image can be recognized according to its themes distribution. Experimental results show that our method can perform better than the variational inference algorithm over IDOL2 database and our own image set captured under different imaging conditions in our campus.