Visual map summarization based on the distance histogram of the landmark

Abstract

The essential function of self-driving car is navigation driving car from the current position to the destination. To get to where you want to go, you must first find the exact location of the self-driving car. In general, the Global Positioning System (GPS) is unable to estimate the global position well unless there are no skyscrapers near self-driving car. Visual maps built by the camera are an alternative to these GPS problems. A typical technique is Visual Simultaneous localization and mapping (Visual SLAM), which build a camera map as the vehicle travels and estimates the current position on the visual map. Visual SLAM needs to be consistent across maps, which makes it difficult to manage large maps, such as city-scale or larger. Companies such as Mobileye, TomTom, and HERE are dealing with large-scale maps by introducing cloud-servers to solve map size problems. Exchanging visual maps between the cloud-server and the vehicle is via the mobile network. With the advent of 5G, data transfer of up to 10Gbit/s is possible, but since the amount of data is directly involved with the cost, the smaller the size of the visual map to exchange, the better. The vehicle sends a mobile map to the cloud-server, which manage mobile maps to build the server map used for estimating the location of the vehicle. Since almost more than 90% of the map cloud-servers and the vehicle exchange consist of landmarks, which are three-dimensional points, it is necessary to reduce the amount of map exchanged on the mobile network by effectively reducing landmarks. In this paper, we propose the distance histogram that represents the distance distribution between a vehicle and a landmark , and propose a new map summarization method that improves location estimation accuracy at high summarization rate with proposing distance score obtained from the distance histogram. Exchanging visual maps between the cloud-server and the vehicle is via the mobile network. With the advent of 5G, data transfer of up to 10Gbit/s is possible, but since the amount of data is directly involved with the cost, the smaller the size of the visual map to exchange, the better. The vehicle sends a mobile map to the cloud-server, which manage mobile maps to build the server map used for estimating the location of the vehicle. Since almost more than 90% of the map cloud-servers and the vehicle exchange consist of landmarks, which are three-dimensional points, it is necessary to reduce the amount of map exchanged on the mobile network by effectively reducing landmarks. In this paper, we propose the distance histogram that represents the distance distribution between a vehicle and a landmark , and propose a new map summarization method that improves location estimation accuracy at high summarization rate with proposing distance score obtained from the distance histogram.