Ren Xiaofeng, chief scientist of Alibaba Gaode map: high precision algorithm promotes the landing of high-precision map


The cloud habitat 2020 conference was held online from September 17 to 18. Alibaba Gaode map, together with its partners, has carefully organized a “smart travel” special session to share with you the importance of building a smart travel platformDT+AIandComprehensive Cloud ArchitectureUnder theNew generation travel and life service platformDuring the process of thinking and practice, we mainly shared topics such as “high-precision map, high-precision algorithm, intelligent space-time prediction model, automatic driving, AR navigation, Lane level technology”.

This is the first article of “godtech” which has compiled and published the main contents shared by the lecturers.

Ren Xiaofeng, chief scientist of Alibaba Gaode map, shared the topic of “high actuarial method promoting the landing of high-precision map”. Ren Xiaofeng introduced the challenges of high-precision map making and landing from the perspective of algorithm, and how to polish and break through key technologies to make high-precision map industry leader.

Last year, Ren Xiaofeng gave an overall introduction to the application of visual technology in Gaode, mainly sharing the work on conventional maps, AR navigation and visual positioning. This year, the time is relatively short (each lecturer has only 12 minutes to give a speech), and mainly introduces the work of high actuarial method.

What is a high-precision map? It’s mainly a map for autonomous driving scenes. It’s very different from the map we usually use when driving. One key is to be precise. A high-precision map needs accurate elements, precise location and fine representation. For example, the poles, lane lines and driving tracks in the figure below need to be accurately expressed in a high-precision map.

Ren Xiaofeng, chief scientist of Alibaba Gaode map: high precision algorithm promotes the landing of high-precision map

In order to land successfully, high-precision map is a must. High precision maps have many uses: positioning for autonomous driving; modeling “recognition” of static targets; and path / behavior planning. These are very important functions for autonomous driving.

What are the main challenges of high-precision map landing?

Precision! Precision! Precision!(important to say three times) the other challenges are scale, cost, and timeliness.

The first is precision. Gao De’s first principle for high-precision map making is to be accurate! To achieve high accuracy, 100 meters relative accuracy of 10 cm. In the real world, if there are two elements 100 meters apart in the real world, their relative position must be within 10 cm compared with the real world. This is a very high goal. Under such a high goal, production efficiency and cost have become very big challenges. The specific details will not be explained here. At present, Gaode has achieved the leading level in the industry. In addition, there are freshness and update. At the earliest time, it was based on the year, and then it was updated monthly, weekly, daily and even faster. Only in this way can the map information be accurate and effective.

The algorithm plays an important role in solving the challenges of high-precision map. There are three main parts of the work

  • Data accuracy and alignment
  • Identification and production automation
  • Change discovery and update

Data accuracy

The first principle of high-precision maps is accuracy. Gaud has invested a lot from the beginning, using expensive collection vehicles. The high-precision RIEGL acquisition vehicle, high-precision laser radiance, its ranging accuracy can reach 5mm, 1m / sec; high-precision integrated inertial navigation; thousand search base station solution But even if such a large investment and acquisition vehicle equipped, it does not mean that the collected data will be no problem, there will still be problems. For example, trajectory, static or motion will make errors. There are not many errors, which may be only 0.5%. However, errors will cause problems in point cloud data, such as point cloud layering. Engineers need to do a lot of algorithmic work to detect and repair the point cloud when it is layered.

After a series of work, the track error rate can be significantly reduced. After the problem of single data acquisition accuracy is solved, the problem of multiple data alignment will be encountered soon. For example, it is possible to collect multiple points on a road. If the point cloud is not processed, it will appear very obvious ghost. It is not pasted together, and it needs to be aligned with algorithm. This also has a great challenge, because the accuracy requirement is very high, within 5cm.

There are many challenges in various scenarios, such as the impact of vegetation. This happens when data is collected in different seasons, and these trees and shrubs can have a very big impact. At the same time, it is necessary to keep the track rigid during alignment. Because when the original acquisition comes back, the relative accuracy of the trajectory is very good. The relative accuracy cannot be destroyed or reduced during alignment. However, it is necessary to improve the relative accuracy on the original basis, including the up and down scenes. If it is a road, the data collected from the two directions and observation angles of the up and down directions, as well as the bridge on and off the bridge, the alignment challenge is higher because of the limited common view area.

Data alignment

How to do data alignment? It is divided into two parts: front end and back end.

There is a core algorithm in the front endPoint cloud matching。 More common, such as ICP or GICP algorithm. However, it is not enough to pay attention to point cloud matching. To solve this problem well, many other problems need to be solved. Such as sparse point cloud feature extraction, fast point cloud semantic segmentation, fast lane line segmentation. These two aspects are related to semantics, and they are a big challenge in efficiency. Because the amount of data in point cloud is very large, we can’t spend too much time on computing time.

The front end has done a lot of work, so what does it do to align the back end,The back end is mainly for large-scale optimization。 Because track trimming can not be done in a single point, it needs many tracks, even in the size of the whole city to make adjustments together.

Golder also spent a lot of time doing optimization algorithms. For example, we do a sparse optimization algorithm based on spline curve. If we make it sparse, we can achieve a hundred fold acceleration. It can better solve the point cloud alignment problem in a city scale.

Identification and production automation

After the alignment problem is solved, it is the production efficiency, that is, the problem of identification and production automation. There are many elements in high-precision maps. For example, linear elements include lane lines, guardrails, curbs, natural boundaries, and the so-called obj, namely, ground signs, poles, traffic signs, bridges, gantry, etc. These need to be made out, using automatic method, algorithm method is a very important link. Its input isPoint cloudandimage recognition。 HD map elements can be generated by algorithms. Algorithms can be used to improve labor efficiency.

Take a few challenging examples. For example, there are four situations in the figure below:

Ren Xiaofeng, chief scientist of Alibaba Gaode map: high precision algorithm promotes the landing of high-precision map

How to solve it?

The input is mainly point cloud and image. Golder has spent a long time in optimizing model capabilities and improving accuracy. include:

  • Semantic segmentation of point cloud
  • Image panoramic segmentation (detection / segmentation depth fusion)
  • Point cloud / image fusion (front fusion + post fusion)
  • Traditional algorithm aided (e.g. fitting, 3D)
  • Vectorization and modeling (depth feature + graph model)

Gaode has achieved a very high level in these methods: high accuracy recall > 98%; partial realization of skip operation, reaching > 99.5% recall; partial realization of no manual inspection, reaching > 99.5 accuracy.

Change discovery and update

In terms of map updating, golde has two sets of plans, which are applied to different scenes. One is the laser method, the other is the visual method.

Laser method。 In order to control the cost, a relatively low-cost laser and a relatively low-cost integrated inertial navigation system are used. The quality of the input data is relatively low, and many things need to be done to improve the accuracy of the data. include:

  • Tightly coupled lidar slam / lio
  • Real time semantic segmentation
  • Location maps: multiple feature layers
  • In the loop relocation
  • Covariance model
  • Global pose optimization
  • Positioning layer update

These are relatively mature solutions. Although the input data is poor, it can still achieve good accuracy in the update scene.

Visual methods。 It is made of very low cost consumer camera and very low cost integrated inertial navigation system. There are many visual algorithms in the process. include:

  • Tightly coupled visual slam / Vio
  • Location feature layer and semantic layer
  • Feature + semantic reorientation
  • Global pose optimization (fusion of vio / relocation)
  • Location feature layer and update

Now it has reached 15cm; under the condition of strict evaluation, it is already the leading level in the industry. The accuracy will continue to be improved in the future.

Visual updating technology can be directly applied to map construction. The image below shows the overlap of visual mapping and satellite images.

Ren Xiaofeng, chief scientist of Alibaba Gaode map: high precision algorithm promotes the landing of high-precision map

How can high-precision maps achieve higher accuracy and higher frequency of updates?

Suppose there is a very low-cost scheme to collect images. The collected images must be of poor quality. On the basis of these images, we need to compare them with the changes in the real world. For example, in the two pictures below, there is an electronic eye, which is actually the same electronic eye, but it looks very different on the picture. We need to use the method of algorithm to judge whether it is the same or not Electronic eye. There are a lot of image algorithms to do. include:

Ren Xiaofeng, chief scientist of Alibaba Gaode map: high precision algorithm promotes the landing of high-precision map

Gao de has been doing all these things, and has achieved good results. This is a necessary process in the process of high-precision map making. This is also the characteristic of Gaud, because with these low-cost devices, we can use the existing low-cost information to discover changes in the physical world.

High precision map is an important direction of Gaode in the future. Its production and landing is a systematic project, in addition to the algorithm, there are many other key work to be done. Let’s work together.

Ren Xiaofeng, chief scientist of Alibaba Gaode map: high precision algorithm promotes the landing of high-precision map