“China Standard” HD Map Structure

There are 3 commonly adopted data structures for HD map on the market right now that are NDS, Local Dynamic Map and Here HD Live Map. In China, a unique structure is initiated by China Industry Innovation Alliance for the Intelligent and Connected Vehicles (CAICV). In this article, I illustrate this “China standard” and the relevant technologies CAICV is researching.

HD map players nowadays are facing the same challenge, which is to build in scale & update in time at lower cost. Eventually another challenge to auto OEMs and Tier 1s in their L4 & L5 vehicles would be to integrate the HD map so that the map’s safety goal is maximized in vehicle perception, planning and control automatically without human interference.

This challenge is less discussed because we are at the stage of L2 & L3 where the map is a reference to human and human driver still plays a key role operating the vehicle. Ultimately on the automation landscape, human wouldn’t be in the middle and it is the machine doing direct planning and control based on information from map and other sensors.

It is not easy to integrate HD map into the autonomous driving system partially because everyone is collecting and organizing map data in their own way. The market is in need of a standard data structure for all the players to follow. The following 3 are quite common at this point:

• NDS.Classic/NDS.Live proposed by NDS Association (Read details here https://nds-association.org/)
• Local Dynamic Map proposed by DMP (Read details here https://www.dynamic-maps.co.jp/en/service/index.html)
• Here HD Live Map (Read details here https://www.here.com/products/automotive/hd-maps)

CAICV, short for China Industry Innovation Alliance for the Intelligent and Connected Vehicles, is launched in June 2017 with the mission to promote intelligent and connected vehicles technology research, build the industry ecosystem, and define industry standards and protocols. Within CAICV, the group for Autonomous Driving Map and Localization has nearly 100 members including HD map players, autonomous driving tech companies, OEMs and so on.

CAICV Map and Localization group initiated a 7-layer data structure for HD map based on data attributes as well as how the autonomous driving system is using the map data.

Layer 1: classical road network data, corresponding to static mission planning in the autonomous driving algorism. This layer generates rough routes for vehicles.

Layer 2: dynamic traffic data like congestion, construction and weather, corresponding to dynamic mission planning. Layer 1 and 2 combined are like the navigation app that everyone has been enjoying on mobile phones.

Layer 3: road-lane topological link data, corresponding to lane level route planning in the autonomous driving algorism.

Layer 4: high definition lane data like lane geometries, road rules and POI, corresponding to reference trajectory planning. The first 4 layers combined assuming no obstacles on the road generate a trajectory route for vehicles.

Layer 5: high definition feature data like lane width, lampposts height, corresponding to supplementary/enhanced localization and perception in the autonomous driving algorism.

Layer 6: real-time perception data like traffic lights, other vehicles, bikes and pedestrians from other vehicles’ sensors/v2x, corresponding to dynamic trajectory planning in the autonomous driving algorism.

Layer 7: knowledge data from AI and deep learning for when map isn’t available.

As the layers move up, data volume and precision level are also going up. Layer 1 and 2 are traditional map data of the minimum volume. Layer 5 and 6 are point cloud data, and layer 7 is AI models where the accuracy is down to 5cm.

In this structure, each layer corresponds to a function in the autonomous driving algorism, allowing selected layers of data to be applied as needed. For example, when a vehicle is heading from A to B, initially the vehicle calls for layer 1 and 2 for route planning, followed by calling lane data (layer 3 and 4) only on the defined routes. This design would avoid unnecessary high volume of data transmission in multiple scenarios.

Eventually, HD map can be defined as a “sensor” with historical and real time data on the cloud beyond the vehicle’s field of view that provides perception data together with cameras, radars and LiDARs for cross check, planning and vehicle control.

Besides the HD map structure, CAICV is looking into the technologies to achieve localization and trajectory planning.

When it comes to localization, a common solution on the market is to compare point cloud data from vehicle LiDAR with HD map. Huge volume of data is involved here making it very difficult to scale. CAICV’s research is on using landmarks, which is relatively light weighted. Horizontal lane data and vertical road signs and lampposts are both identical to locate the vehicle.

Reference trajectory planning could be done in 2 steps like the example given earlier. First step is road level. Second step is lane level where the algorism checks for lane details on the defined road level that minimized data involved and optimized processing time.

Further more, CAICV is doing research on dynamic trajectory planning with v2x technologies. With shared sensor data from all vehicles, an individual can definitely see beyond and ahead. The technological challenge here with data fusion is coordinate system and aggregating different sources. HD map makes it easier as a unified platform for data cross check and updates.

The contents above are from a presentation by Jiang Kun from CAICV on Aug 15, 2019. We will see how the HD map structure evolves.

(Initially published on LinkedIn on August 26, 2019.)