Bayesian Estimation for 3D-point Object Identification Based on Probabilistic Field Representation

New network services using real spatial information are expected to emerge in remote areas. For the advancement of services, it is important to understand real space from real spatial information, and for this purpose, object identification using machine learning is widely employed. Instead of directly identifying real space information using machine learning techniques, in this study, we represented real space as a field of probabilistic superposition of objects, which incorporates the results of the object identification as well as information about neighboring objects. We obtained the neighbor information based on positional relationships of objects in real space from a dataset used by the machine learning algorithm, and built an empirical knowledge base of the positional relationships. Then, we developed a method to use the empirical knowledge to modify the object identification results. Our results show that the outcomes of object identification are changed by the empirical knowledge and the accuracy of the identification is improved when confidence in the machine learning algorithm is low.