Before talking about the high-precision map itself, let's review the basic logic of autonomous driving.
Simply put, there are three main steps to realize automatic driving: perception, decision-making planning and driving control. This is similar to the logic of walking to work/school: your eyes see the picture and tell your brain, then you know where you are and which direction you are going, and command your legs to take a step.
In the automatic driving system, perception mainly solves two problems: what is around the car and where the car is. The car has no eyes. We should give it "eyes". This is its environmental awareness system, which generally consists of visual sensors (cameras), radar (millimeter waves, ultrasonic waves, lasers) and other sensors.
With the cooperation of these sensors, the perception system can know what is around the vehicle, such as cars, people, roads, trees, walls, road signs and so on. After providing these data to the decision-making system, the decision-making system will know whether the vehicle can run, what the maximum speed can be, whether it is necessary to control the front wheel to turn, and transmit it to the traffic control system.
Under certain conditions, the sensor system will directly transmit data to the traffic control system, which is mainly used in AEB, that is, emergency active safety system, to ensure that it can respond to emergencies in the shortest time.
But there is a problem here. You know the way because you are familiar with this route, have a map in your head and know the next step. However, the vehicle can't know how to exercise next only by the perception system, just like you are in a strange city.
That's when you need a location and a map.
In the automatic driving system, vehicle positioning is the key. It affects almost all the links. Through GPS (satellite positioning), IMU (inertial measurement unit) and wheel log, the vehicle can clearly know its position, current motion state (speed, acceleration) and so on.
At this time, the perception system not only knows what is around the car, but also knows which road the car is on and which direction it is going. The decision-making system also knows the road information of the whole area, and will plan a reasonable route according to the destination and give traffic control instructions.
This is a complete automatic driving process. In an ideal state, these are enough for the vehicle to take you anywhere safely.
But unfortunately, with the current technical level, this "ideal state" is difficult to achieve, because the vehicle perception and positioning system can not identify all the information on the way out and judge its accurate position like people, and it is not affected by the environment.
For example, in the city center with many tall buildings, the GPS signal may be blocked, and then the vehicle will lose its position information, thus interrupting the automatic driving journey.
Or, in rainy and snowy weather, the lane line on the road is covered with snow or accumulated water, and it is difficult for vehicles to distinguish lanes only by environmental perception system, which may lead to accident symptoms.
In other words, the current perception system is still difficult to identify potholes, speed bumps, low shoulders and so on. If you miss these things during high-speed driving, it will also lead to more serious consequences.
Or, when going up and down the ramp at high speed, sometimes there will be left and right lanes. At this time, if only the navigation map and environmental awareness are used, the vehicle may change lanes quickly and the riding experience is poor.
In this technical condition, in order to realize automatic driving above L3 level, high-precision maps are needed.
As the name implies, a high-precision map is a very high-precision map. Generally speaking, the accuracy of high-precision maps is decimeter level, but it is not only high in accuracy, but also richer in data dimensions than ordinary maps.
High-precision maps store a lot of driving assistance information in the form of structured data, one of which is road data, such as lane information such as the position, type, width, slope and curvature of lane lines. The other is the information of fixed objects around the lane, such as traffic signs and traffic lights, road details such as lane height limit, sewers and obstacles, and infrastructure information such as overhead objects, fences, numbers, road edge types and roadside landmarks.
Image source: Future car lecture hall-Netease cloud classroom high-precision map, the only way for automatic driving
Simply put, with the help of high-precision maps, vehicles can already know the road direction, curvature, detailed lanes, speed limit, how far the road sideline is from the shoulder, how many trees, fire hydrants, telephone poles, where there are speed limit signs, where there are significant landmarks and so on.
And this information makes the autopilot system directly open the "god mode".
With the help of high-precision maps, the vehicle positioning can be more accurate. Even in areas with complex road conditions, the positioning system can feed back more accurate lane information to the decision-making system, which can plan lanes/routes.
At the same time, because the high-precision map contains a large number of static reference objects, the positioning system can compare the environmental information obtained by the environmental awareness system with the map information, thus calculating the actual position of the vehicle in the case of poor GPS signal, and improving the robustness of the whole system.
For the perception system, a high-precision map can define the area that needs to be recognized. Showing this area to the perception system is the focus of your image analysis, and reducing the recognition of other areas, which is the region of interest (ROI). Using ROI can reduce the load of perception system, liberate computing power and increase the recognition accuracy of key areas.
For example, under normal circumstances, the front camera only needs to focus on the lower part of the picture, because the sky is above the picture, and vehicles and people will not appear from the sky. But traffic lights are an exception. They are often hung high, so if you want to recognize the actual traffic lights, you have to keep searching from the whole picture, which puts a lot of pressure on computing power. However, if the ROI is turned on, a signal light will appear in a certain area of the screen at a certain point in the high-precision map, then the system can search for the signal light and complete the identification as long as it focuses on identifying the marked area.
For decision-making system, with the help of high-precision map, the complexity of decision-making algorithm can be reduced, as long as the vehicle can drive to the destination as smoothly as possible without collision. Because there are detailed information about lanes and fixed obstacles in the high-precision map, as long as you follow the planned route, you will definitely not hit the pit, ride on the shoulder, or even hit the telephone pole.
At the same time, detailed lane information can also make the system plan a more reasonable and stable driving path. For example, the ramp problem mentioned above, if the system already knows the existence of the fork in the road in advance, then it will change to the corresponding lane in advance before seeing the road to improve driving comfort.
Moreover, because the lane division, lane width and speed limit of each lane are marked in detail in the high-precision map, the system can plan more efficient routes and the decision-making system can focus more on driving safety during driving.
Generally speaking, vehicles without high-precision maps will have a "conditioned reflex" reaction when driving automatically: they only know how to slow down when they see traffic lights; I didn't know I couldn't hit the pole until I recognized it. With the help of high-precision map, the vehicle is ready to slow down before driving to the traffic lights; The escape route was planned in advance before touching the telephone pole.
It can be said that high-precision maps make self-driving cars "prepared".
But at present, there are some shortcomings in high-precision maps. First of all, because of its large amount of information, it will be more difficult to collect, and the collection cycle will be lengthened accordingly. Therefore, the current high-precision map can not cover all roads, basically only highways and major urban roads.
At the same time, in the continuous construction of the city, the elements marked by high-precision maps may change, which puts forward high requirements for the real-time update of high-precision maps, which is also a major difficulty in the application of high-precision maps in urban road conditions.
Seeing this, you will find that the high-precision map is actually a "dispensable" thing: if the recognition success rate of the automatic driving system is high enough, you only need the plane navigation map commonly used by human beings, and you can also realize fully automatic driving.
Tesla, for example, has always insisted that high-precision maps are not needed because they think their recognition and decision-making algorithms are strong enough. In the near future, they only need to rely on real-time recognition and processing (conditioned reflex) to complete fully automatic driving on the basis of ordinary maps.
Unfortunately, at present, most manufacturers, including these independent manufacturers who are about to launch L3 production cars, are not as good as Tesla in the accumulation of autonomous driving algorithms. In order to make up for the lack of perception, we need the over-the-horizon perception provided by high-precision maps and a large number of prior information supplements.
With the help of high-precision maps, there is no problem to realize L3 automatic driving under the national standard. This also explains why high-precision maps have been widely used since autonomous driving entered L3 era.
It not only publicizes its own technical strength, but also can be used as a selling point and gimmick. Why not? What do you think?
This article comes from car home, the author of the car manufacturer, and does not represent car home's position.