A New Semantic Descriptor for Data Association in Semantic SLAM

To successfully implement SLAM based on semantic information, object recognition is essential. In our semantic SLAM approach, the robot should localize itself knowing only the pose of the surrounding objects. The implicit information in TOSM contains conceptual knowledge, in which we stored data that cannot be perceived by sensors only. However, it needs to distinguish not only an object class, but also which specific object instance a given detected object is related to. In many approaches, objects are recognized by detecting feature points and representing them as descriptors. There are several types of descriptors for feature points, such as SIFT and SURF. There are also descriptors describing a whole object instead of just feature points, like GOOD or HOOD. We suggest a new semantic descriptor, which includes more high-level information, and propose a process to recognize accurately through semantic analysis by semantic descriptors. In this paper, we introduce about our semantic descriptor model while giving some example cases, and then describe the process of data association by using the aforementioned descriptor.