[Abstract] Autonomous driving has become the strategic development direction of the global automobile industry. Among them, L3 highway autonomous driving is the most likely autonomous driving system to be implemented first. High-precision maps and positioning systems are a key part of the autonomous driving system. They have developed in recent years. Quickly, it has reached mass production status. The article first analyzes the development status of autonomous driving and high-precision map positioning. Then, it studies the application of high-precision maps and positioning systems in geofence determination and perception redundancy of autonomous driving systems. Finally, it initially proposes a high-precision map Positioning test plan.
【Abstract】Autonomous driving has become the strategic development direction of the global
automotive industry． Among them，L3 autonomous driving on highways is the most likely to be implemented first． The high － precision map and positioning system are key parts of the autonomous driving system，which develop rapidly and reach a state of mass production． First，the development status of autonomous driving and high － precision map positioning is analyzed． Then，the application of high －precision maps and positioning systems to the determination of geographic fences and perceived redundancy of autonomous driving systems is studied． Finally，a test program for high － precision maps and positioning systems is proposed．
【关键词】高精度地图高精度定位高速公路自动驾驶
0 Introduction
In recent years, with the continuous increase in car ownership, what follows are: (1) Serious traffic congestion, low traffic efficiency, and increasing time spent on
traffic ; (2) Frequent traffic accidents, traffic The number of casualties and costs caused by accidents are increasing, and most
accidents are caused by human factors; (3) Air pollution is becoming increasingly serious. In addition, the aging population is serious and will soon become a major problem facing the world. The travel problems of the elderly and other people who cannot drive cars need to be solved urgently. As an important method to solve the above problems
, autonomous driving has become the strategic development direction of the global automobile industry.
In addition to major car companies that regard autonomous driving as their core development area, many auto parts suppliers, Internet companies and startups are also focusing on developing autonomous driving-related businesses. High-precision maps and positioning systems are a key sensory input for autonomous driving, which can provide prior information on the road thousands of meters ahead, including information on roads, lanes, traffic signs, and road ancillary facilities.
1. Development status of autonomous driving and high-precision map positioning systems
1. 1 Development status of autonomous driving systems
The automatic driving system includes L3-L5 level driving automation systems (based on the International Society of Automatic Mechanical Engineers SAE J 3016 [1] and automobile driving automation classification [2]). As shown in Table 1, during the activation of the automatic driving system, target and event detection The system is responsible for both response and control of the vehicle's lateral and longitudinal movements. The autonomous driving system consists of an autonomous driving domain controller, a sensing subsystem, a control execution subsystem, a human-computer interaction subsystem, and a power supply subsystem. The sensing subsystem is a
key and is responsible for the perception of the system. It generally consists of monocular cameras, binocular cameras, night vision cameras, millimeter wave radar, lidar, ultrasonic radar, high-precision maps and positioning, It is composed of V2X and other components, and the appropriate sensor combination and detailed parameters of each sensor are selected according to the sensing requirements of the system.
Because highway working conditions are more uniform and standardized than urban working conditions, they are closed scenarios. The driving direction of vehicles is single, and there is hard isolation from opposite driving vehicles. There are fewer types of vehicles allowed to drive, and the slope and curvature of the road, lane width, etc. There are unified requirements in laws and regulations, so L3 highway autonomous driving is currently one of the main research and development directions of various car companies. At the end of 2017, Audi announced that TJP (Traffic JamPilot, speed range 0-60 km/h) for L3 highway autonomous driving in traffic congestion scenarios has reached mass production status. Mercedes-Benz and BMW plan to launch full-speed L3 highway autonomous driving in 2021. L3 level autonomous driving requires that when the system issues a takeover request, the driver needs to complete the takeover within a certain period of time. At the same time, it also needs to ensure that it can only be turned on within the geofence. There are risks in turning it on outside the geofence. High-precision maps and positioning systems are the best way to ensure that autonomous driving can only be turned on within geo-fences.
This article analyzes the application of high-precision maps and positioning in autonomous driving on L3 highways.
Another research direction of autonomous driving is L4 autonomous driving in suburban working conditions and fixed areas. L4 level autonomous driving means that the driver does not need to take over dynamic driving tasks. The system can cope with situations beyond the system's designed operating range, system failure, etc., and requires higher redundancy design of the system. Suburban working conditions are more complex than high-speed working conditions, but the speed range is lower. Companies such as Google Waymo, Baidu, General Motors and Ford are all focusing on developing L4 autonomous driving. By the end of 2019, Waymo's actual vehicle test mileage has reached 20 million miles (32 million kilometers), and Waymo's takeover rate in 2019 was 0 ．076, that is, manual intervention is required every 13,219 miles (21,150 kilometers). High-precision maps and positioning are also necessary components of L4 autonomous driving, and this article does not conduct application analysis.
1. 2. Development status of high-precision map positioning system
1. 2. 1 The development status of high-precision maps
The United States has long begun the layout and research and development of high-precision maps. In addition to traditional navigation companies such as Mapbox, there are also companies such as Waymo, Ushr, General Motors, Ford, Civil Maps, and DeepMap. Dynamic Map Planning, an investment fund jointly established by the Japanese government and private industry funds, Japan Innovation Network Corporation, Mitsubishi Electric, Toyota and other companies jointly established Dynamic Map Planning to promote the industrialization of high-precision maps. Here and TomTom companies in Europe, as well as domestic companies such as AutoNavi, Baidu, and NavInfo, are also early in the field of high-precision maps [3]. AutoNavi has mass-produced high-precision maps in 2018, and Baidu’s high-precision map products will be mass-produced in 2020.
Currently, L3 autonomous driving generally uses vector high-precision maps, and L4 autonomous driving generally uses high-precision maps in both point cloud and vector formats.
The vector high-precision map used by L3 autonomous driving has more elements than ordinary navigation maps and ADAS maps, and its relative accuracy and absolute accuracy are higher.
In terms of map element composition, the navigation map contains road-level road network information and POI information for users. ADAS maps contain information such as road-level road network, slope, curvature, speed limits and number of lanes. In addition to ADAS maps, high-precision maps also have lane-level road networks, detailed lane models (shape points or geometric information of all lane lines and curbs/guardrails, lane-level curvature, slope, speed limit, height limit, etc.) and user interfaces. Based on high-precision positioning feature information, there will also be lane-level real-time dynamic information in the future.
In terms of relative accuracy and absolute accuracy, the current industry level of high-precision maps used in mass production of L3 autonomous driving is an absolute accuracy of
1 m (2 Sigma) and a relative accuracy of 20 cm (2 Sigma).
1. 2. 2. Development status of high-precision positioning.
Ordinary GNSS has large absolute positioning result errors due to satellite ephemeris errors, satellite clock errors, ionospheric refraction, tropospheric refraction, multipath effects and receiving equipment errors, generally in the range of 2-10 m. Currently, real-time dynamic carrier phase difference technology (RTK) is commonly used to correct the above errors. RTK broadcasts the carrier phase observations collected by the base station to the vehicle based on the vehicle position. The vehicle calculates high-precision absolute position information through the RTK solution algorithm. The absolute positioning accuracy in open scenes can generally reach 1 m (2 Sigma). If positioning is only based on GNSS/RTK and high-precision maps, there will be road positioning errors in some scenarios (such as parallel service roads next to highways). Therefore, it is necessary to combine feature matching and positioning methods [4] to improve the accuracy of positioning. Currently, among the smart driving sensors on mass-produced vehicles, the front-facing camera can output detection information. The front-facing camera can output 4 lane lines (the left and right lane lines of the current lane, the left and right lane lines of the side lane) and traffic signs. and other information, by combining the lane line type (dashed line, solid line, dashed solid line, solid dashed line, etc.), geometric information (lane width, lane line heading angle and curvature, etc.) and color (white, yellow, etc.) with high-precision map data Match to determine the vehicle's location. However, cameras are easily affected by factors such as lighting and vehicle occlusion. Subsequently, it is necessary to combine detection information from other sensors such as lidar and millimeter-wave radar to improve the robustness of positioning.
2 High-precision map positioning technology and its application in autonomous driving systems
2. 1 Key technologies for high-precision maps and positioning solutions
L3 level autonomous driving adopts a high-precision positioning solution based on multi-sensor fusion. High-precision positioning is performed based on the feature information of multi-sensor information fusion. The feature information includes lane line attributes (geometric parameters, type and color, etc.), lane attributes (width, etc.), traffic signs, traffic poles, guardrails, curbs, etc. (see figure 1), the characteristic information will be detected by sensors such as cameras, millimeter-wave radar, and lidar.
Absolute position information is obtained through GNSS/RTK. Absolute position related information includes positioning status, positioning quality, number of satellites currently used for positioning, longitude, latitude, speed, accuracy factor, etc. Absolute position information and trajectory estimation information based on IMU, wheel speed, steering wheel angle, etc. are fused to obtain absolute positioning results. Based on the absolute positioning results, a certain range of high-precision map data is extracted and matched with the feature information detected by the vehicle-mounted sensor. After the feature matching is completed, the vehicle's road-level positioning (which road), lane-level positioning (which lane) and in-lane positioning are determined. (Lateral position, longitudinal position and heading angle) results. Because it adopts a multi-sensor fusion (camera, millimeter wave radar, lidar, IMU, GNSS/RTK, etc.) solution, when a certain sensor becomes invalid within a certain period of time (for example, the camera's lane line detection confidence is exceeded due to strong light). The system can still maintain high-precision positioning results when GNSS/RTK has no valid data due to low or excessive occlusion. The multi-sensor fusion algorithm is the key to high-precision positioning based on the information of each sensor. The fusion algorithm needs to balance the differences of each sensor in different environments (different input information confidence levels). Therefore, the robustness of the high-precision positioning solution based on multi-sensor fusion is and better reliability. The fusion algorithm has the characteristics of multiple changes. For different sensor configurations, loose coupling or tight coupling can be used to fuse sensor data at different levels. Multi-sensor fusion also has a variety of mathematical methods to choose from, such as Kalman filtering, Particle filtering, multi-Bayes estimation methods, etc.
One of the key points is that the autonomous driving domain controller and the high-precision map positioning system need to be time synchronized. There will be a certain time delay from the time of each sensor environment detection to the high-precision positioning system receiving the characteristic information. Time synchronization solutions include hard synchronization and Soft synchronization solution: Hard synchronization generally uses hard wires to synchronize two controllers, while soft synchronization uses the Autosar standard protocol to synchronize two controllers.
2. 2 Application of high-precision maps and positioning in L3 highway automatic driving systems The application of
high-precision maps and positioning in L3 highway automatic driving systems can be divided into two parts: First, the determination of geofences in the automatic driving system; It provides high-precision map information of the road ahead.
2. 2. 1 Geofence determination
Only high-precision maps contain lane-level attributes, so the determination of lane-level geofences can only be achieved based on high-precision maps and positioning. At the same time, high-precision map positioning can realize the determination of road-level geofences. Road-level and lane-level geofencing are shown in Tables 2 and 3.
(1) Geofence 1:
Different areas and roads in areas not open to autonomous driving may have special scenarios. Before mass production of autonomous driving, large-scale, sufficient mileage roads and simulation verification are required, so the autonomous driving system is more suitable in different areas. Or the roads are gradually opened. As shown in Figure 2, autonomous driving can be restricted to area A and not allowed to be activated in area B. It can also be restricted to some sections of road A and autonomous driving can be activated in other areas of road A. Autonomous driving is not allowed on the road section.
(2) Geofence 2: The area where the road is about to end.
The autonomous driving system needs to ensure that the driver takes over or stops safely before the end of the highway. The high-precision map positioning
system needs to classify the area where the road is about to end as outside the geofence, and can provide advance reminders. The driver takes over.
(3) Geofence 3: Road directly connected to the ramp area.
For autonomous driving systems that do not support ramps or automatic off-ramps, because the ramp scene is different from the main highway road, for example, the
curvature of the ramp is greater, etc., some roads will directly connect to the ramp ( The main road of the expressway disappears). The autonomous driving system needs to ensure that the driver takes over or stops safely before entering the ramp. The high-precision map positioning system needs to classify the area directly connected to the ramp as outside the geo-fence, which can remind the driver to take over in advance
. Similarly, geofence 11 is the area directly connected to the ramp of the current lane. High-precision map positioning needs to classify this area as outside the geofence.
(4) Geofence 4: Tunnel area
The tunnel scene is more complex. For example, changes in light entering and exiting the tunnel will have a certain impact on perception, and the risk of failure is greater in the tunnel. Therefore, some OEMs classify tunnels as outside the geofence. The autonomous driving system needs to ensure that the driver takes over or stops safely before entering the tunnel. The high-precision map positioning system needs to demarcate the tunnel area as outside the geo-fence to remind the driver to take over in advance.
(5) Geofence 5: Toll station area
The lane lines in front of the toll station are generally irregular, and other vehicles change lanes more aggressively, so some OEMs classify the toll station area as outside the geofence. The autonomous driving system needs to ensure that the driver takes over or stops safely before entering the toll station area. The high-precision map positioning system needs to classify the toll station area as outside the geo-fence to remind the driver to take over in advance
.
(6) Geofence 6: Road construction area
The road construction scene is complex and the lane lines are irregular. The autonomous driving system needs to ensure that the driver takes over or stops safely before entering the road construction area. The
high-precision map positioning system needs to divide the road construction area into For areas outside the geo-fence, the driver can be reminded in advance
to take over. Similarly, geofence 12 means that the construction area is in front of the current lane. The current lane is under construction and cannot be passed. The high-precision map positioning system needs to classify the lane construction area as outside the geofence.
(7) Geofence 7:
When the road lane lines are missing in the area where the road lane lines are missing, the vehicle will drive irregularly. The autonomous driving system needs to ensure that the driver takes over or stops safely before entering the area where the road lane lines are missing. The high-precision map positioning system needs Demarcating areas with missing road lane lines as outside geofences can alert drivers in advance to take over. In the same way, geofence 13 is the area with missing lane lines at the lane level. The current lane has the problem of missing lane lines. The high-precision map positioning system needs to classify the area with missing lane lines at the lane level as outside the geofence.
(8) Geofencing 8: Area with missing road guardrail.
When the left road guardrail is missing, the probability of pedestrians or animals entering the highway is greater. When the right road guardrail is missing, the risk is greater because there is no hard separation from the opposite lane. The autonomous driving system needs to ensure that the driver takes over or stops safely before entering the area where the road guardrail is missing. The high-precision map positioning system needs to classify the area where the road guardrail is missing as outside the geo-fence, which can remind the driver to take over in advance.
(9) Geofence 9: Emergency Lane Area
Under normal circumstances, vehicles are not allowed to drive in the emergency lane, and the autonomous driving system needs to comply with traffic regulations. The autonomous driving system needs to ensure that it is not allowed to be turned on in the emergency lane, and the high-precision map positioning system needs to classify the emergency lane area as outside the geo-fence.
(10) Geofence 10: Lane disappearance area
When lanes merge, as shown in Figure 4, three lanes merge into two lanes, and the rightmost lane disappears. The autonomous driving system needs to remind the driver to change lanes before the lane disappears or automatically complete the lane change. The high-precision map positioning system needs to mark the lane disappearance area as outside the geo-fence, which can alert the driver in advance.
(11) Geofence 14: The
autonomous driving system in areas with excessively wide lanes needs to ensure that the driver takes over or stops safely before entering the area with excessively wide lanes. The high-precision map positioning system needs to classify the area with excessively wide lanes as outside the geofence. This can be done in advance. Alert the driver to take over. In the same way, the high-precision map positioning system needs to classify the area where the lane is too narrow (geofence 15) as outside the geofence.
(12) Geofence 16: Area with excessive lane curvature
When the lane curvature is too large, autonomous driving lateral control may deviate from the lane and other risks. The autonomous driving system needs to ensure that the driver takes over or safely stops before entering the area with excessive lane curvature. The high-precision map positioning system needs to locate the area with excessive lane curvature. Zoned outside the geofence, drivers can be alerted in advance to take over.
(13) Geofence 17: Area with excessive lane slope
When the lane gradient is too large, there may be risks in perception and longitudinal control. The autonomous driving system needs to ensure that the driver takes over or stops safely before entering the area with excessive lane gradient. The high-precision map positioning system needs to divide the area with excessive lane gradient into geographic areas. Outside the fence, the driver can be reminded in advance to take over.
2. 2. 2. High-precision map information for all lanes of the road ahead.
The high-precision map and positioning system will send map data to the autonomous driving system as required, as shown in Figure 5. The distance of the map data ahead and the number of branch roads can be defined according to requirements, such as The sending distance is 1. 5 km, the branch road data is 2 (based on the current location, the two closest roads to the car), and the high-precision map positioning system will send data of PATH1, PATH2 and PATH33 roads. The autonomous driving system can obtain information about the road ahead, lanes and road facilities in advance based on the road ahead data provided by high-precision map positioning, which plays an important role in reducing erroneous perceptions and making behavioral decisions in advance.
The high-precision map elements are shown in Table 4. The high-precision map positioning system can also be regarded as a sensor that is not affected by the outside world. Even if some vehicle-mounted sensing performance declines or fails, it can still provide information such as lane lines ahead within a certain period of time, which can play a role in sensing redundancy. Typical scenarios include: (1) When an autonomous vehicle encounters bad weather such as rain, snow, fog, etc., the detection performance of the on-board sensors (camera detection performance is poor, lidar detection performance will also be affected to a certain extent) will be affected; (2) When the front-view sensor such as the forward-view camera is blind or malfunctions; (3) The front-view sensor such as the forward-view camera is blocked by other vehicles and affects the detection of lane lines and other information.
2. 3 High-precision map and positioning system test
High-precision positioning test content includes road-level positioning, lane-level positioning, lateral error and longitudinal error of in-lane positioning. The true value of longitudinal positioning is tested through high-precision absolute position measurement equipment, such as Novatel SPAN CPT The POS320 of Maipu Spacetime calculates the longitudinal error based on the true value and high-precision positioning results, and other indicators are evaluated through the manual annotation results based on the true value camera. The high-precision map test is performed by storing and comparing the true value information of the sent signal.