Relevance Vector Machine based multi-feature integration for semantic place recogntion

In order to work in realistic scenarios, it is a desirable feature for autonomous robots to extract semantic concepts from environments. In this paper, A Relevance Vector Machine (RVM) based approach is presented for the task of visual semantic place recognition. The high sparsity and Bayesian property makes this approach capable of obtaining probabilistic confidence estimation, and computationally efficient during the online prediction stage. Meanwhile, in order to take advantage of discriminative powers of different feature descriptors, a multiple kernel technique is introduced in our system, resulting in a very flexible model where arbitrary feature descriptors can be integrated smoothly. In this paper we choose three popular descriptors for our implementation. Experiments carried out on real typical office environment datasets show the feasibility and robustness of our approach.