Map and perception fusion to help automatic driving
Beijing, China, May 24, 2019—In the “China Automated Driving HD Map Technology Innovation Conference”, Mr. Li Bijun, Professor of State Key Laboratory of Surveying, Mapping and Remote Sensing Information Engineering of Wuhan University, participated in the conference and made a keynote speech. With the theme of “Map and perception fusion to help automatic driving”
First of all, Professor Li introduced what is a map. The map is a map of the natural and social phenomena on the surface of the earth, which is selectively drawn on the plan paper, according to certain mathematical rules and with prescribed schema symbols and colors. Traditional electronic navigation maps are maps that are stored and viewed digitally using computer technology. The high-precision map in the automatic driving is a static vehicle road network including high-precision, road markings, road peripheral facilities, and road dynamic information updated in real time. The current autopilot maps should all be a scene map, based on the safe behavior, dynamically correlate various information on the road, and comprehensively reflect the digital refinement map of the location itself and the various features, events or things related to the road. . Due to the local perspective of the scene map, rich content, dynamic changes of the target, complex geometric relationships, and difficulty in continuous registration. So add awareness to complete the automatic driving.
The purpose of perception is to identify lane lines around the car, road access areas, vehicles, pedestrians in real time. Traffic signs, etc. How to perceive these contents and how to ensure its reliability is a big challenge at present, because there is still a big gap between artificial intelligence technology and human intelligence. No matter how the artificial intelligence algorithm develops, there are still many limitations.
Professor Li then explained to us the development of maps in the automotive field and the need for maps for autonomous driving. The development of electronic maps for vehicles has gone through three stages: the first stage is the basic navigation electronic map, which aims to assist the driver in navigation and provides basic road navigation functions; the second stage is the ADAS level map, which aims to achieve active safety. ACC, LDW, LKA, FCW and other functions are available; the third stage is a high-precision map of HAD level, which is designed to achieve automatic driving, providing high-precision positioning, road-level and lane-level planning, and lane-level Guidance ability, etc.
Because the map can provide prior knowledge, you can plan your driving control in advance, you can know where I am, what is going around me, what is going on in the future, and how to go where I want to go, so I need a map; because of advanced assisted driving or Automated driving requires precise intelligent control of the machine and requires accurate environmental basic information, so automatic driving requires high-precision maps.
High-precision maps have lane-level accuracy maps (eg, lane geometry and road geometry; lane attributes, traffic signs, traffic lights, intersections, etc.; lanes and inter-element attachment and association; lane-level topology; including dynamic and static obstacles (.) and characteristics of real-time traffic information. With high-precision maps, automatic driving will quickly land, and high-precision maps are the only way to drive automatically. High-precision maps enable corner warnings, speed limit consultations, frequent accident road warnings, hazardous area warnings, slope-based fuel savings, adaptive headlights, and fatigued driver monitoring. It can solve the problem of partial perception and greatly reduce the difficulty of perception.
Since the electronic map updating technology is a complex system engineering involving many problems such as collection, fusion, compilation, differentiation, transmission and integration of map information, the physical storage format of electronic map data is the basis for solving this problem. However, there are many types of physical storage formats for navigation products, the compact structure is unfavorable, and there is no uniform standard. There will be a series of problems such as production links, update links, and imperfect laws and data specifications. Then Professor Li provided us with the following four solutions: 1. From professional measuring vehicles to low-cost collecting vehicles to crowdsourcing update mode. 2. Develop data content of high-precision maps as early as possible. Updated standards for data organization management and transmission. 3. High-precision and high-reliability high-precision map data processing algorithm. 4, V2X, road coordination.
Next, Professor Li explained how to achieve autopilot through map-assisted perceptual positioning: First, build a lane-level high-precision map based on open source navigation electronic maps, providing search space and space constraints. Then, the multi-source observation value is provided in real time through the environment sensing module, and the relative pose of the self-vehicle and environmental information is calculated through real-time CCD data, point cloud data, and GNSS data. Finally, combined with the environment perception observation and high-precision map for real-time positioning and positioning.
Finally, Professor Li believes that maps and perceptions complement each other: high-precision maps provide prior knowledge and global information to reduce perceived difficulty and requirements, and perception provides dynamic real-time information for high-precision maps and even map-assisted positioning. In automatic driving, maps are indispensable, and perception is irreplaceable. The fusion of the two enhances reliability, applicability and reality. There is still a long way to go for autonomous driving.