The battle for map routes in autonomous driving

High-precision map

High-precision maps (HDMaps) are a love-hate relationship in autonomous driving. In recent years, various forces in the industry have also launched heated discussions on whether to use HDMaps. At present, many companies, including Apollo, Weilai, Ideal, Xiaopeng, and Momenta, have used semantic maps (Semantic HDMap).

Compared with ordinary navigation maps SDMap, HDMap has higher accuracy and contains more complete information. SDMap can generally only achieve road-level accuracy and does not provide specific coordinates such as lane lines and road signs; HDMap can generally achieve decimeter-level accuracy, including accurate lane lines, traffic lights, signs, road arrows, and even some high-precision maps indicate the change points of lane line virtual and real attributes, lane line bifurcation points, merging points, etc.

HDMap

The advantages of using HDMap in autonomous driving are obvious:

• Open the third eye: gain the ability to perceive beyond visual range, which is equivalent to opening the God's perspective;

  
  

• High precision: Decimeter-level, or even centimeter-level, precision in certain scenarios can provide very reliable prior information to the positioning and regulation modules, greatly reducing the difficulty of the algorithm and the probability of accidents.

  
  

• High richness: HDMap has a very rich amount of information, generally including roads, lanes, lane lines, road arrows, stop lines, fences, traffic signs, traffic lights and other elements. The map contains the geometry, attributes, colors, topological relationships, etc. of the above map elements. In addition, there may also be lane change points, road curvature, road slope, etc.

Its limitations are also obvious:

• Cost: HD map vendors usually use expensive data collection vehicles to build maps. A single sensor may cost millions, in addition to labor costs. In addition to the map building process, map transmission and maintenance also require certain costs.

  
  

• Coverage: China is big, and the earth is even bigger. HDMap cannot achieve as high coverage as SDMap. Although map vendors such as AutoNavi claim to have covered all highways in China, coverage in urban areas is still a big problem. With the implementation of autonomous driving, application scenarios are gradually moving from highways to urban areas, and roads that are not covered by high-precision maps can only be navigated blindly. This is why many companies limit their autonomous driving in urban areas to a small section, such as Anting in Jiading, Shanghai, and Yizhuang in Beijing, because only these sections have HDMap;

  
  

• Freshness: The situation on the road is changeable, especially in China, where road repairs and cones are commonplace. It is not easy to expect HD maps to respond to these changes in a timely manner. After the map is built for the first time, how to maintain the map continuously is a big problem. A practical problem that many autonomous driving companies need to face is, when the HD map information and perception information are inconsistent, who should they trust? How to judge whether the HD map is not updated in time or the perception is wrong;

  
  

• Regulations: Maps are sensitive information, especially high-precision maps. In China, even SDMaps must be biased (biased by the National Bureau of Surveying and Mapping, Mars coordinate system). There are only a few companies in China with Class A map qualifications, and the country has tightened qualification reviews in recent years. There are not many map vendors that can provide high-precision map services. In addition, China does not allow the sale of maps with height information, which is probably very unfriendly to six-degree-of-freedom posture.

In China, in addition to large map providers such as Tencent, AutoNavi, Baidu, NavInfo, and Zhonghaiting, there are also smaller map providers such as Kuandeng, Deep Motion (which has been acquired by Xiaomi), and Juefei that can provide high-precision map services.

In terms of regulations, the Ministry of Natural Resources has recently released some good news, but the speed of progress is still not too optimistic.

Ministry of Natural Resources-Guidelines for the Construction of a Standard System for Basic Maps for Smart Vehicles (2023)

Crowdsourced maps

With Mobileye's REM as a pioneer, domestic and foreign autonomous driving companies are also promoting the route of crowdsourcing mapping: collecting a large amount of traffic trajectory information or single-vehicle SLAM mapping results, integrating them into accurate semantic maps in the cloud, and using the results of crowdsourcing mapping to update and repair existing HDMaps.

The technical route of crowdsourcing mapping can make up for the shortcomings of the above-mentioned high-precision maps to a large extent: instead of maintaining expensive collection vehicles, low-cost autonomous driving equipment can be used on mass-produced vehicles to make up for quality with quantity.

At present, domestic autonomous driving algorithm companies (Momenta), new forces (NIO, Xiaoli, and Li Auto), chip companies (Horizon Robotics), map providers (Baidu), etc., all have the ability to build maps through crowdsourcing. Some crowdsourcing map building teams targeting mass production have even become quite large in scale.

The RoadMap published by Huawei Qintong Group at ICRA2021 is quite representative: the results of local mapping of multiple vehicles are uploaded to the cloud for fusion and compression to obtain an accurate lightweight semantic map. Subsequent vehicles download the map online and perform 6Dof positioning based on the map.

Huawei Crowdsourcing Solution RoadMap Pipeline

Huawei Crowdsourced Map RoadMap

It sounds great, but crowdsourcing map solutions also face several very realistic challenges: data security, policies and regulations, and user privacy. In particular, data security issues have become more sensitive after the Didi incident, and a red line that cannot be crossed has been vaguely drawn.

In addition, crowdsourcing mapping also faces many technical difficulties that are difficult to handle with long-tail cases.

Navigation Map

Navigation map SDMap is the most common map form at present. It has been popular in our lives for many years. For example, Baidu Maps,

Amap.

The advantages of SDMap compared to HDMap are:

• Low cost of building and maintaining maps

  
  

• Mature development, covering a wide range of areas

  
  

• The regulatory resistance is much smaller than HDMap

  
  

But SDMap has some key shortcomings compared to HDMap:

• Low accuracy: Usually only road-level accuracy is available, but cannot reach lane-level accuracy, making it difficult to meet downstream planning and control requirements.

  
  

• Little information: There is generally no specific geometric information about lane lines and roadside edges, nor accurate location information about traffic lights and signs, which poses a great challenge to positioning and planning control.

As mentioned before, when the high-precision map-based autonomous driving solutions were affected by factors such as regulations and costs, and their advancement was limited, mainstream players turned their attention to SDMap, which is the light map solution we will talk about next.

Light map solution (SDMap+ / HDMap-)

For mass-produced autonomous driving, high-precision maps are still too heavy.

Five years ago, many players, including Xiaopeng and Huawei, ran through highway sections and a small part of urban sections based on high-precision maps and released beautiful demos. However, with the evolution of technology and the acceleration of commercialization, the delivery of urban NOA has become the focus of competition among leading players. The high-precision maps that were once "really good" have become the biggest obstacle to large-scale mass production. Therefore, the light map solution came into being.

It is not rigorous to say that it came into being at the right time, because the light map solution is not new. For example, Musk made it clear long ago that Tesla does not use high-precision maps. The reasons are still the coverage, freshness, and cost mentioned above. (Tesla only uses high-precision maps in the annotation stage, and the FSD actual car operation is only connected to the navigation map)

The light map solution means not using the heavyweight HDMap, but based on the lightweight SDMap, or based on a map format between HDMap and SDMap, combined with the powerful perception capabilities of a single vehicle to achieve autonomous driving.

The light map solution hopes to get rid of the shackles of high-precision maps and reduce the degree of dependence on maps. Correspondingly, the requirements for the perception ability of a single vehicle are also greatly increased. It is precisely because of the progress of visual perception technology in recent years that the "light map, heavy perception" route has become possible, and the Transformer and BEV solutions have made great contributions to this.

In fact, to this day, most players' main projects are still highly dependent on HD maps. After all, without HD maps, it is difficult to make a demo. Without a demo, a company cannot raise money and will not survive the day of light maps. Even if many companies take the route of crowdsourcing maps, strictly speaking, they are still HD maps, but they are produced in a crowdsourcing way.

In 2022, against the backdrop of tightening map regulations and enhanced perception capabilities, light-map and heavy-perception routes have once again become a popular choice.

In September 2022, Haomo shouted the slogan "focus on perception, not mapping" at AI Day