Proposal of Place Detection Method on Topological Map by Combining Global Scene Recognition with CNN and Local Features detection: -Based on survey results of human recognition of outdoor environment by eye tracking-

Generally, an autonomous mobile robot is equipped with various sensors such as a web camera and a laser range finder (LRF). However, it is difficult to read necessary information from a lot of obtained information and process it appropriately. On the other hand, humans can grasp the situation appropriately from surrounding information even in the environment where they have visited. Focusing on this, we investigated the human environment recognition method using a gaze measurement device. As a result, it became clear that humans specified he location from the whole landscape and characteristic objects. In this study, based on this, we propose a point recognition method that combines global scene detection with CNN and local features. As a result, it was found that the landscape recognition using CNN can reduce errors caused by angles and positions, and is more robust than GIST features and superior in scene recognition.