Construction and Localization of Multi Resolution Map Based on Vision

We propose a new visual map, a multi resolution map with cell as the unit, to avoid the various problems that exist in the construction of high precision maps. In the process of human navigation, the ability to understand the scene is very important, and human beings do not need accurate location information. Therefore, the establishment of a non high precision map with the ability to understand the scene is more in line with the habit of human navigation. Based on the image data we collected, we use CNN to extract low dimensional features, and construct cells by clustering low dimensional features. The images and labels in the cell are input into CNN as training data, so that the trained CNN has the ability to locate to the cell. We construct cell map along the road of 2000m in length. When the number of cells is 200, the localization accuracy is more than 98%. When the number of cells is 100, the localization accuracy is more than 99%.