Introduction to high-precision maps for autonomous driving

HD map

---

---

Preface

What is HD map

The main difference between High Definition Map (HD MAP) and ordinary navigation electronic maps is that it has higher accuracy and richer information. The higher accuracy is mainly reflected in the higher accuracy of the absolute coordinates of the high-precision map (referring to the accuracy between a target on the map and the location of external real-world things), which can be accurate to the centimeter level; the richer information is mainly reflected in High-precision maps not only contain road information, but also cover almost all surrounding static information related to traffic.
Compared with ordinary navigation electronic maps, high-precision maps contain richer and more accurate road traffic information. In addition, in terms of application scenarios, ordinary navigation maps are mainly used by drivers, while high-precision maps are machine-oriented and used by self-driving cars.
Accuracy is the biggest difference between high-precision maps and ordinary navigation electronic maps. The accuracy of ordinary vehicle electronic navigation maps is generally about 10 meters. When high-precision maps are used in the field of autonomous driving, they need to be accurately positioned on a specific lane. They also need to know all the surrounding road and traffic information that may participate in autonomous driving decisions. The accuracy requires Reaching 10~20 centimeters, this accuracy is basically the same as the width of a lane edge, so as to ensure that intelligently driven cars will not cross into other lanes and avoid the risk of side collisions with other vehicles.

Ordinary navigation electronic maps should depict roads (links), while high-precision maps should not only depict roads, but also depict the number of lanes (lanes) on a road, truly reflecting the actual style of the road.
The richer information of high-precision maps is mainly reflected in the following aspects:
accurate road shape: slope, curvature, heading, elevation, and roll data of each lane.
Detailed lane line information: whether the lane lines between lanes are dotted lines, solid lines or double yellow lines, the color of the lines, road isolation belts, and the material of the isolation belts will all be described.
In addition, crosswalks, billboards along the road, speed limit signs, traffic lights, roadside phone booths, etc., which are often collectively referred to as LandMark Objects, have absolute geographical coordinates, physical dimensions, and their unique characteristics that will also appear in high-precision data.

High-precision data distribution engine

ADAS (Advanced Driver Assistant System) applications require the road network and attribute data information in front of the vehicle for decision-making control and judgment. Ordinary digital map data is usually only used by the navigation system, but high-precision map data can be used in the vehicle. It is used by other ADAS applications, so it needs to rely on high-precision data and a high-precision data distribution engine for high-precision data broadcast.

ADASIS (ADAS Interface Specification) defines the concept of "ADAS Electronic Horizon", which expresses the road network in front of the vehicle and the road network attribute information. In order to realize this expression, we need to build the vehicle's position model and the possible road models of the road network in front of the vehicle. The accessible roads can be expressed through a tree-like hierarchical structure. In addition, the geometric shape and related attributes of the road will also be expressed by establishing related attribute models. "ADAS Electronic Horizon" data is serialized and transmitted via the vehicle's Ethernet network.

1.3 Definition of terms

1. ADAS(Advanced DriverAssistance System)

That is, the advanced driving assistance system uses on-board sensors to sense the vehicle environment and integrates calculations to allow the driver to detect possible dangers in advance, effectively improving the safety, economy and comfort of vehicle driving.
ADASIS (Advanced DriverAssistance System Interface Specification)
is an industry international standard developed by the ADAS Forum, which is used to standardize the standard interface protocol for exchanging map data between map data and vehicle ADAS applications.
AHP (ADAS Horizon Provider)
is a high-precision data distribution engine that provides beyond-line-of-sight road ahead and data information for ADAS applications.
AHR (ADAS Horizon Reconstructor)
is used to parse the messages sent by AHP and reconstruct map data for use by the terminal ADAS application module.

2. Why is a high-precision data distribution engine needed?

The high-precision data distribution engine serves as a bridge between high-precision data and ADAS applications. Its value can be summarized in the following aspects: the
need for long-range sight distance for autonomous driving. As a map sensor for autonomous driving, high-precision maps can provide more reliable over-the-horizon vision. distance range to support more reliable decision-making judgments. The need to improve accuracy and the transition from guides to guided vehicles have increased the accuracy requirements. Interface standardization for high-precision map data distribution.

Construction of high-precision data distribution engine

The relationship between high-precision data distribution engine and ADAS applications

The data distribution engine involves the following components and interactions:
AHP
AHR
ADASIS V3 Protocol
ADAS application, see the terminal application department in the figure above for details

High-precision data distribution engine architecture

The high-precision data distribution engine consists of multiple layers, including the engine layer, protocol organization layer, and system adaptation layer. The relevant platform and tool support are shown in the figure below: Engine layer: loading, analysis and lane network data of high-precision

data organization.
Protocol layer: Mainly assembles protocol messages with data provided by the engine layer and delivers them to the adaptation layer.
Adaptation layer: Mainly responsible for docking and interacting with the system, and distributing the organization's protocol data to ADAS applications.

Model expression of high-precision data distribution engine

Abstraction and expression of road network model

The road network model of the data distribution engine includes three layers of model abstraction. First, the real world model is abstracted into a high-precision road network model, and then the high-precision road network model is further organized and divided into a tree model expressed by Path and Offset.
Expression of the abstract model of the real world

Digital map model and navigation path set by the user, expression of map elements Expression of the

vehicle position and road network in the digital map model Expression of the road
network model near the vehicle position through links to express the relationship between the road network connection relationship. In digital map databases, a road network is represented as a set of connections and nodes that define the links between them.

From the perspective of ADAS applications, the road network behind the vehicle is not concerned, so the data distribution engine is composed of the road network in front of the vehicle.

The road network in front of the vehicle is organized by Path. Each Path is a set of links. The road network data in front of the car can be expressed by two algorithms.
In the simple Path method, starting from the link where the vehicle is located, each passable path is independently expressed as a Path. The
optimized path organization method reduces data redundancy and can also fully express the road network data in front of the vehicle.
Therefore, the data distribution engine describes the shape of the road network in front of the vehicle and its surrounding environment as a collection of different path and map data attributes to form a prediction tree. This prediction tree is connected by multiple paths, each path represents a part of the road and the intersection between roads.
Once a vehicle moves and changes its position, the predicted view also changes, some paths behind the vehicle may be removed, or new ones in front of the vehicle may be added. The characteristics of the path are expressed as a set of attributes, such as the number of lanes, geometric shape, curvature, etc. included in the expressway and urban expressway network itself. The position of an attribute on a path is represented by a set of offset values, which are distance markers that define the absolute distance along the path itself, expressed in centimeters. The origin of a path is the zero offset value point, and the offset value of an attribute represents the distance between the attribute itself and the origin of the path. If the path is newly started and has no parent path, the offset value 0 is the starting position of the vehicle.

Attribute model of high-precision data distribution engine

The attribute model data of the data distribution engine comes from the attribute information on the high-precision road network, which is defined as expressed along the Path and defined as the position on the Path, expressed through Offset. For example, the Speed ​​Limit attribute provides speed limit values ​​for points on a path.
The attribute model can be divided into the following three different types according to the interpolation type, namely Spot, Step, and Linear types.

Spot type attributes are only valid at a given Offset position in Path, and the different attributes are expressed by different Offset positions. For example, a traffic light can be defined as a Spot type attribute because it can be expressed as the existence of this point attribute at a certain position in Path. A
Step type attribute is defined to be valid until the Offset position of the next attribute. The attribute is expressed as a value in the range from Offset to EndOffset on Path.

In the example shown above, the Path length is 200. Speed ​​limit 80 is valid overall, from Offset 0 to 200. There are two speed limit values ​​starting from offset 50 and 100. Therefore, the attribute distribution on the entire graph is as follows:

Offset 0: Starting speed limit value 80. Offset 50: The rainy day speed limit value of 60 is introduced, and the speed limit of 80 continues. Offset 100: Repeat speed limit 80, add fog limit 50, end with speed limit 60 in rain. Offset 150: Repeat speed limit 80, end with speed limit 50 in fog.

Properties of type Linear are defined as linear differences between given positions.

Linear interpolation attributes are not expressed continuously. At the same offset, the value on the left and the value on the right are different. The attribute model uses the following method to express such discontinuous attribute values.
At Offset, store an attribute, the value stores the attribute value on the left, and EndOffset is 0. Stores a property at the same Offset and the value stores the property value on the right, but EndOffset > Offset

Car location information model

In the data distribution engine, the location information of the car can be expressed through Path and Offset. In the case of uncertainty, the car's location may exist on multiple Paths, so a set is needed to describe the car's location information. The following information can be expressed through the car position information:
whether the car information is out of the data area. Whether the car information matches the data range of Path. Whether the car information is matched to multiple Paths. Whether the vehicle information enters or leaves the data area.
The TimeStamp value of the vehicle position information expresses the time value at which the sensor information is received.

The vehicle's position information can also express the more likely path ahead.

As shown in the picture above, the possible path on the left is P1, and in the picture on the right is P3.

Synchronization mechanism between high-precision data distribution engine and receiving end

The data distribution engine synchronizes the road network Path data between AHP and AHR through pathControl messages.
When the pathControl message does not contain a certain Path, AHR deletes the Path in the road network after receiving the message. When the pathControl message remains unchanged from the last time, AHR maintains the current road network unchanged after receiving the message. When a pathControl message adds a certain Path, AHR adds the Path information after receiving the message
and synchronizes the attribute data through profileControl.

The interaction mechanism between the high-precision data distribution engine and the receiving end

The data distribution engine (AHP) and the receiving end (AHR) have the following interaction mechanisms:
broadcast mode request/provide mode subscription/publish mode.

The current high-precision data distribution engine uses the "request/provide" method. By sending ADAS messages, AHR can request and feedback information.

Assist AHP and ADAS application integration

Main AHP and Auxiliary AHP

Not all data in the ADASIS protocol is provided by the data distribution engine, and an auxiliary AHP engine can also be added. Auxiliary AHP engines can send sensor information or sensor fusion information.

The main data distribution engine and the auxiliary AHP engine are formed.

Two integration methods of ADAS applications

According to the main AHP and auxiliary AHP engines, two types of ADAS application integration methods can be realized, namely downstream integration and upstream integration.

Downstream integration

There is no fusion processing on the AHP side. Each sensor data and high-precision map data are transmitted to the AHR side through communication for fusion processing, and then passed to the ADAS function application.

Upstream fusion
performs fusion processing on the AHP side, and passes the fusion results to AHR for processing through the protocol, which directly affects the ADAS function.

Quality construction

In order to ensure software quality, the following technical means are used in the construction of high-precision data distribution engines:
unit testing function testing quality inspection tools
visualization tools
visualization tool screenshots

Typical architecture application form

According to the high-precision data distribution engine architecture, it can be divided into the following integration forms:

The data distribution engine (ie EHP engine) is integrated into the map box

The concept of map box
is a car-grade integrated software and hardware product that carries "map data + high-precision positioning" capabilities, which is different from pure software products. (The following names all refer to the high-precision map box: Positioning Box/MAP ECU/MAP BOX/HDLM...among them L:Localization M:Module)
including content
maps and related applications: HD data, AHP, positioning, OTA...Basic software: system, Bottom-level driver, diagnosis...Basic hardware: system-on-chip (SoC), memory, storage, IMU (optional), protective shell...Network and communication interface: CAN/Ethernet input, Ethernet output, USB interface...Solution features and clear division
of
labor : Car companies can use this architecture to dismantle functions into small modules and put forward product requirements for management and control respectively to avoid being unable to start with a completely black box solution. Suppliers can be replaced when delivery risks are encountered.
Functional safety considerations: Details such as chip selection, hardware design, network security, and system diagnosis can be left to professional suppliers; there are uncertainties in functional safety levels such as map quality, online updates, and backhauling, and require cooperation with AD ECUs. Isolated to enable the AD ECU to meet functional safety requirements.
Facilitates the management of high-end and low-end products: you can choose products with different configurations from suppliers.
Reduce the computing power burden on domain controllers: Make it easier to find functional safety hardware that meets computing power requirements.

Integrated in IHU

Solution features
: Cost reduction: No additional hardware modules are required.
Integrate the V2 solution and reduce uncertainty: AHP V2 is mostly on the car side, and the solution has been run through, so using a similar method for maps and V3 can avoid the uncertainty of the new architecture.
It is easy to advance due to internal reasons of car companies: For some car companies, if the high-precision map business planning and navigation map department promotes the box solution from the bottom up, it will cause major changes to the overall structure and it will be difficult to advance.

Integrated within domain controller

Features of the solution
: Reduce the occupation of vehicle network bandwidth by cross-domain communication: Most of the sensors used for sensing are connected to domain controllers. If the map and positioning are placed on the domain controller, back-end applications can be used directly or indirectly without cross-domain communication. Maps to reduce the usage of in-vehicle network bandwidth.
More suitable for car companies that take the self-research route and choose an overall solution: For car companies that take the self-research route and choose a single solution provider to provide a complete solution, there is no need to deploy functional modules separately.

Scenario application examples

High-precision positioning applications

Combined with high-precision data to assist horizontal positioning and vertical positioning.
Vertical positioning is mostly combined with relevant information such as road signs obj and lane geometry, while horizontal positioning is mostly combined with relevant information such as lane lines and guardrails.
Active safety applications often combine sensor (millimeter wave radar, camera) information and map data for matching and correction, thereby improving positioning accuracy.

High-speed autonomous driving (HWP)

The driving environment for function activation
mainly relies on map judgment: (1) Expressway and city express; (2) Clear lane lines; (3) Curvature and slope; (4) No objects or events that trigger alarms or braking: including dynamic road environments; (5) ) is not at night and the weather conditions are good (visibility is more than 200 meters).
Implementation functions
Take cruise lateral control in this lane and autonomous parking in this lane in abnormal scenarios as examples:
Lane type: Autonomous driving relies on lane type to divide the drivable area. If the type is wrong, the vehicle will drive in a non-driving area, which will bring consequences to the self-driving vehicle. Potential safety hazards; at the same time, in the scenario of autonomous safe parking, if the lane type is wrong, it will directly lead to the loss of autonomy and safety of autonomous safe parking. Lane line type: The auxiliary camera performs lane line recognition; compares it with the camera to perform lane keeping.

Automatic cruise based on navigation route

Function activation
: The working environment depends on the map judgment:
road grade: highway/city block. Whether PartOfcalculateRoute (navigation path identification) is continuous without interruption. Weather type: Sunny/light rain/cloudy weather conditions allow the function to be activated.
Implementation function:
When getting on/off the JCT, it will determine whether to get on/off the JCT based on the navigation path mark and the road network in front of the vehicle, and will remind the side of the lane change ramp in advance. When automatically changing lanes into JCT/merging into a highway, the auxiliary camera will recognize the lane line according to the lane line, and compare it with the camera to detect and compare the line shape to determine the timing of the vehicle's lane change.

future evolution

On the one hand, consider further integrating the architectural design of AHP V2 and V3 to better assist autonomous driving. In addition, as part of the data closed loop, data provision and recycling capabilities are enriched.
Reference link