Several major issues regarding high-precision positioning for autonomous driving

However, an engineer responsible for positioning at a new power car company said that it is completely "unacceptable" for them to leave visual fusion positioning to positioning box suppliers. The reason is that, on the one hand, it is “related to who integrates whom,” and responsibilities are not easy to distinguish. “If there is a problem, is it a problem with the sensor, the positioning box, or the fusion algorithm? It is difficult to determine the blame.” ; On the other hand, they do not recognize the image processing capabilities of the positioning box supplier. Moreover, this also involves the waste of computing power caused by the image being processed twice.

Regarding this behavior, an executive of an integrated navigation supplier believes that at this stage, when traditional car companies or car companies with weak R&D capabilities are mass-producing high-end intelligent driving, in order to more conveniently control costs and improve reliability, they prefer to supply Providers can provide overall positioning solutions, such as integrated satellite navigation, inertial navigation, high-precision maps, and even visual fusion positioning and other functions.

The solution of integrating combined navigation into domain control has high requirements on the research and development capabilities of car manufacturers on the one hand, and on the other hand it also has high requirements on algorithms such as perception and decision-making. Therefore, if neither of these two requirements can be met, leaving the positioning to a third-party supplier is a solution that can achieve rapid mass production with relatively high integration and reliability.

It seems that although the integration of positioning modules is a trend in the long term, positioning boxes may still exist for a long time.

eight. Completing the fusion positioning algorithm can reduce the requirements for key components of integrated navigation

The integrated navigation mentioned above is actually part of the autonomous driving positioning system. The final system locates by integrating various data.

△Pony.ai integrated positioning architecture

(Source: Pony.ai account Zhihu article)

Commonly used data for fused positioning include GNSS, IMU, RTK (or PPP-RTK), high-precision maps, wheel speed sensors, lidar, cameras, etc. The final positioning accuracy depends on the positioning accuracy after fusion.

△Apollo integrates high-positioning software architecture

(Data source: Apollo official)

It is worth mentioning that in addition to absolute positioning based on GNSS, lateral relative positioning can also be achieved based on the position of the lane line observed by the camera, as well as surrounding key points observed using lidar, cameras, millimeter-wave radar, etc. Features (such as signs, traffic lights, etc.) are matched with high-precision maps to achieve vertical relative positioning (SLAM).

△Relative positioning matching principle

(Source: Liufen Technology’s online sharing)

In fact, if the fusion positioning algorithm is done well enough, it can reduce the requirements for key components of integrated navigation (such as IMU accuracy, etc.).

Many experts have said that the high accuracy of IMU is mainly to deal with some positioning degradation scenarios-such as tunnels. In the tunnel scenario, absolute positioning can only be performed by relying on the GNSS values ​​at both ends, and the IMU can only be relied on for track estimation in the middle of the tunnel.

However, it is not impossible. An expert from Baidu Apollo mentioned in a sharing: On the one hand, the requirements for absolute accuracy in the tunnel are not high, and the curvature changes of the road will not be too extreme; on the other hand, more precise methods can be introduced. Many features, such as textures, can help with positioning.

Some experts also mentioned that vision and millimeter-wave radar can be used to achieve lateral positioning in tunnels. As for the accuracy of longitudinal positioning, "it is actually not that important."

Nine. Can integrated positioning get rid of dependence on high-precision maps?

In addition to reducing the requirements for integrated navigation, a company that has entered L4 pre-installation mass production also revealed that it is exploring solutions to achieve high-speed NOA functions without relying on RTK.

In fact, high-precision positioning is used with high-precision maps. If RTK is not needed, is high-precision map still needed?

Many people in the industry believe that high-precision maps are a "crutch" in autonomous driving. High-precision maps are not affected by the environment or limited by distance, and can assist perception for over-the-horizon sensing input, effectively complementing the sensor. If you throw away the "crutch" now, there may still be some problems. For example, in real-world roads, you often encounter problems such as large cars blocking the road and blurred lane lines. Sensors are often limited by distance and environmental influences.

According to feedback from some industry experts, the lack of high-precision maps (relying only on navigation maps) makes it difficult to achieve lane-level positioning. Because the navigation map lacks lane information on the road, relative positioning can only be performed through lane information recognized by vision. When encountering road signs and lane line obstructions at intersections, it is difficult to identify the road topology, and decision-making and planning errors are prone to occur.

In addition, when driving on a ramp with a small curvature, because the navigation map lacks curvature information, the vehicle's horizontal and vertical control will be more difficult, and it is easy to have a jagged feeling.

From these aspects, some degree of "high-precision map" is still needed, but this so-called "high-precision map" may not be as accurate or as complete as the elements provided by current map dealers. It may be a navigation map. It is overlaid with a layer of road semantic map, which can be called "Navigation Map Plus" for the time being.

In this regard, memory parking provides a good paradigm. Before the memory parking function is turned on, you need to build a map and memorize the route first, and then make positioning decision-making and planning based on the constructed map.

Therefore, with reference to memory parking, some solution providers have proposed the memory driving function. For urban commuting routes, the map and route memory are constructed after repeated driving several times, that is, a layer of locally constructed semantic map "Navigation Map Plus" is overlaid on the navigation map, and then positioning decisions and planning are made based on this map to realize the urban commuting route. Point-to-point navigation assistance that does not rely on high-precision maps.

This can achieve a "stand-alone" version of the map to a certain extent. What if we go one step further and share the locally built map to the cloud? Wouldn’t that just become a crowdsourced map?

In fact, Xpeng Motors said that it already supports the parking lot memory parking map sharing function. After users learn the route, they can choose to share the route. After passing the cloud review, they can manage and share the route on the mobile phone to reach the target parking lot. , you can also use parking routes uploaded by other users or officially recommended. To a certain extent, this is already a de facto "crowdsourcing" map of the parking lot.

Although driving and parking have different functional logics, they may reach the same goal in solving map problems.