A new navigation method allows autonomous vehicles to navigate without detailed maps and without road markings present. Autonomous navigation thus also becomes possible in rural areas, for example. The technique was developed by MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL).
Researchers at CSAIL write in a research paper that autonomous driving can have a significant positive social impact, but its application is significantly limited by the fact that autonomous vehicles struggle to navigate in rural areas. Autonomous driving systems currently generally rely on two navigation methods:
- navigate by following road markings;
- navigate based on highly detailed and accurate maps.
Navigation in rural areas
Both methods are less suitable for navigation in rural areas. For instance, in such areas, road markings are in many cases missing or of such poor quality that autonomous vehicles have difficulty registering the markings correctly. In addition, in many cases it is not practical to develop detailed maps of rural areas. This is because many of these areas are very vast and, in addition, the environment is constantly changing due to the presence of a lot of vegetation.
MIT has therefore developed a new navigation technique that allows autonomous cars to navigate without detailed maps of their surroundings. The technique combines simple maps with basic information with information gathered from sensors to find its way safely. In this process, the vehicle chooses a navigation target within the range of its sensors, after which the navigation system chooses a route based on the available simple map. Sensors then measure, among other things, the vehicle's distance from the side of the road to ensure the vehicle stays safely on its own road side. These sensors also continuously search for obstacles, such as people, animals, vehicles and other objects. This process is continuously repeated until the vehicle reaches its final destination.
OpenStreetMap
The technique uses OpenStreetMap, a project that aims to develop freely available and freely editable maps. These maps are created by volunteers, who collect and store information about streets, rivers, borders and areas, among others, in a public database. OpenStreetMap also contains information on traffic rules on roads. MIT's navigation technology uses GPS to determine the location of the vehicle on these maps. By combining this information with data collected by sensors fitted on the vehicle, the vehicle can find its way in areas where detailed maps are not available.
The researchers point out that their technique additionally makes it possible to store a map of a much larger area on the same amount of storage space. For example, the researchers point to Jesse Levinson, Michael Montemerlo and Sebastian Thurn of the Stanford Artificial Intelligence Laboratory, who described how they managed to compress a detailed 2D map of over 32,000 kilometres of road into a data file of about 200GB in the 2007 article 'Map-Based Precision Vehicle Localisation in Urban Environments' in Robotics: Science and Systems. By using a map with much less detail, a map of this same area can be reduced to just 3.5GB.
Speeds of up to 108 kilometres per hour
The technology was tested by MIT during a full-scale experiment, in which an autonomous Toyota Prius managed to find its way around a rural area using the system. This vehicle is equipped with LIDAR sensors to detect the distance to objects and inertial measurement systems to measure a vehicle's speed and direction of travel, among other things. The data collected by these sensors is translated using software running on a standard PC. Using the system, the vehicle can detect the road up to 35 metres ahead.
The researcher claims that this is sufficient to allow a vehicle to safely drive autonomously at a speed of 30 metres per second, equivalent to about 108 kilometres per hour. By running the software on a graphics processor in parallel, the researchers expect that this maximum speed can be increased substantially.
More information can be found in the researcher's paper from MIT's CSAIL.
Author: Wouter Hoeffnagel
Source: MIT
Source photo: CSAIL from MIT
