How difficult is it to make a high-precision map? Listen to the big dream of start-up companies.

Original title: how difficult is it to make a high-precision map? Listen to the big dream of start-up companies.

How difficult is it to make a high-precision map? Listen to the big dream of start-up companies.

Liu Jun, founder and CEO of wide stool Technology

There is a key point in making high-precision maps: collecting and making maps is only 10% of the work done, and 90% of the work will be updated in the future. Crowdsourcing is an important means of updating maps.

With the announcement of the unmanned “armament race” by the auto factories, the leading auto factories have completed the commercialization of auxiliary driving (L1), semi-automatic driving (L2) has begun to be used on the strings, and conditional automatic driving (L3/L4) or even the future of full automatic driving (L5) is the unattended driving commanding heights that everyone must contend for. .

However, the most important part of unmanned driving is the matching of high precision maps. The so-called high-precision map, compared with the ordinary navigation map used by our mobile phone at present, the latter navigation and positioning relies on GPS, the accuracy error is about 10-15 meters, contains a very single map elements, obviously this kind of low-precision, less elements of the map is unable to meet the unmanned needs of vehicles.

What is the use of high-precision maps?

High-precision maps, as the name implies, are quite accurate, with errors of less than 15-20 centimeters, and can be detailed marked out the specific lanes on the highway, lane width and length, street lamp location, you can understand that a real-world road is fully vectorized into a map.

With high-precision maps, when the vehicle is unmanned, it will really “understand the world,” know the width of the parallel line, know where the lane line is, should not parallel line. With the help of the camera, lidar and so on, the vehicle can really be driverless.

How difficult is it to make a high-precision map? Listen to the big dream of start-up companies.

Examples of high precision maps

High-precision maps are a track where traditional big companies like Google, Baidu, Gold and Four Dimensions are investing, and innovative companies are catching up. Broadstool Technology is a technology company focused on making high-precision maps. Its founder and CEO is Liu Jun, former chairman of Baidu take-out and vice president of Baidu LBS.

It is only a year since the founding of Broadstool Technology, announced in February this year that it has received hundreds of millions of dollars of financing from IDG-led investment rounds. Kuantong Technologies is a brand-new start-up company, doing something quite focused, Liu Jun told Titanium Media, Kuantong Technologies only focus on the production of high-precision maps on this matter, and high-precision maps upstream and downstream related to other things, will not be involved, so that we can focus on doing a good job. And it will also bring more opportunities for other upstream and downstream companies to cooperate, which is the origin of the company name “bench”.

How to create a high-precision map?

Domestic companies that really do high-precision maps have only started to do it from scratch in recent years, so everyone’s starting point is basically the same, compared to the methodology and speed of implementation of maps, as well as the financial support behind.

Liu Jun told the Titanium media that there are two main types of high-precision map data acquisition: a, laser radar (ranging) + camera (identifying objects) + GPS; b, camera (identifying objects) + GPS + algorithm (calculating distance).

Theoretically speaking, laser ranging is accurate, but because of the error of multi-sensor fusion, there is no domestic company to solve the problem of laser multi-sensor fusion error, and the core technology of wide-bench technology is to use algorithms to replace the ranging function of laser radar. Through the AI training of a large number of maps, the computer can identify the lane, extract the feature points, construct the lane network topology, and mark the crosswalk, marking, traffic signs and other markers. And these tasks have been highly automated, basically do not need manual map marking, synchronization and other work. After comparing with the actual road, the accuracy can be controlled within 20 cm.

How difficult is it to make a high-precision map? Listen to the big dream of start-up companies.

Camera +GPS data collector for wide stool Technology

So, the wide stool technology is using the camera + GPS + algorithm, obviously this way on the algorithm is more demanding. Liu Jun showed the Titanium media his own camera acquisition device, a 5-megapixel camera, which can be easily mounted on top of the vehicle to facilitate data acquisition. At present, more than 20 vehicles collect data on the road every day, while Beijing, Chongqing, Shanghai and other large cities have basically completed the data collection, it is expected to complete the first high-precision map of the national highway transportation network within the year.

Vehicles on the road with video cameras record GPS latitude and longitude data while running on the road uploading the collected data to the clouds. The data uploaded to the clouds are identified by AI, which identifies different objects and gives them relative latitude and longitude signals. The amount of data collected in a day can reach 600-800G, so the acquisition is not difficult, the difficulty lies in the calculation and identification of the large amount of data after processing.

How difficult is it to make a high-precision map? Listen to the big dream of start-up companies.

Mapping and comparing data of tunnels

After data processing is completed and a high-precision map is formed, it is necessary to manually compare the accuracy of the computer-computed line over the captured video. Finally, 3D high-precision maps with road shapes, object attributes, color labels and latitude and longitude information are obtained. The error of high-precision map is small, and it can do the mosaic of multiple maps, and form a larger national map by networking.

For example, when GPS is lost in a tunnel, AI intelligent identification is needed to ensure the map error; and the gas station beside the highway can not be driven automatically, which needs to be marked on the map, seemingly simple things in fact, after the arrival of automatic driving, is a matter of life. The high accuracy map significantly improves the safety of autopilot.

How can the technology of wide stool be profitable and what is the crowdsourcing plan?

During the interview with Titanium Media, Liu Jun described in detail his high-precision map collection crowdsourcing program: “There is a very important point in making high-precision maps: collecting and making maps is only 10% of the work, the other 90% is in the future updates, and using crowdsourcing is an important means of updating maps.”

The first acquisition of high-precision maps is to use their own teams to run to measure, but the maintenance and updating of the future also rely on their own teams to run, is not realistic, for personnel and capital requirements will be quite high. Then some companies take the form of outsourcing, the renewal of this business to other companies to do, then the capital investment will always have, and relatively large.

How difficult is it to make a high-precision map? Listen to the big dream of start-up companies.

High precision map is the foundation of pilotless driving.

In the form of crowdsourcing, the Broadstool technology works with the automobile factory. In the future, automobile vehicles will be equipped with cameras. If these vehicles are included in the crowdsourcing system, the cost of mapping will be greatly reduced. As the vehicle runs along the road, it constantly returns the changed information to the cloud, and when a place is uploaded many times and has changed, the map of the place will be updated.

Broadstool Technology has already cooperated with many automobile factories. It is expected that the L3-class automobile will be launched by 2020, and crowdsourcing will also start.

It is obvious that Liu Jun had planned to sell high-precision maps to various mainframe factories, and that updating high-precision maps in the future will be profitable in every vehicle sold with unmanned technology.

Liu Jun told Titanium media that although high-precision maps do not seem to have many competitors on the “track”, the technical ability is particularly demanding, and the high-precision map market is actually a “trillion” market capitalization, wide bench technology since its inception, has been running in this direction, without changing the goal. Last month, Kwan Bench Technology officially announced the launch of the “100-million-city plan”, which plans to complete the mapping of 100 cities and 1 million kilometers of roads in the next year, and the completion of China’s main road network map.

“Titanium media to our company interview time is particularly good, it is just that we have been working hard for half a year, conquer technology, produce results can be said, the end of the year to see our first high-precision map products,” Liu Jun said to the Titanium media on the occasion of the interview farewell. Well, let’s look at how this start-up company realizes “big dreams” step by step. (this article starts with titanium media, author / item Europe).

Editor in chief:

AiShangHai LoveShangHai