There is no fixed data collection vehicle for Gaoshen Zhitu, which is a “light asset model”.
The traditional navigation map is for humans to read, including the POI guiding action from here to the other place, while the high-precision map is for machine reading and is an important infrastructure for L3-L5-level automatic driving.
Recently contactedHigh-precision map and location service providerDeepMap (aka “Shenzhen Zhitu”) was established in Silicon Valley, USA in 2016, At the end of 2018, the company was registered in China and established offices in Beijing and Guangzhou.
The company completed Series B financing in early 2019. The investors included Goldman Sachs Group, Nvidia, Bosch Ventures, Generation Investment Management, and in 2017, the company received Andreessen Horowitz, Accel Partners and Jinsha Jiangchuang. Investment in the A round of financing. Gao Shenzhi, general manager of China, Liu Yuquan, told them that their total financing has exceeded 90 million US dollars.
At present, the high-precision map of the high-sense map is mainly for L4-level autopilot, and also considers L3. Its business model includes: providing high-precision map customized production services for traditional car companies, auto-driving startups and other graphic companies that cannot break through technical difficulties; providing positioning services; multi-sensor calibration to achieve coordinate conversion; providing low-cost Data collection program.
Talking about the difference between the service host factory and the autopilot startup company, Liu Yuquan said that the main engine factory cares about the cost and the car rules. Every day, it wants mass production, and wants to gradually evolve from ADAS to L4. Autopilot startups basically start with L4.
FIG advanced chi Zou Liang director of engineering, said , high-precision map production technology are mainly two: First, the laser point cloud integration technology, is to use laser radar scanning, distribution points in the scene to return art for L4 and L5 level autopilot; second is based on an image and GPS technology for ADAS .
where the laser point cloud integration technologies, including two categories, one point cloud integration algorithm based on its wide application scenarios, not limited to the GPS scene, which is the advanced technology used in Figure intellectual ideas; the second is based on a comparison Accurate differential GPS and precision inertial navigation (IMU, inertial measurement unit), but this technology relies heavily on the scene and must be used in open areas. GPS-weak scenes such as viaducts and underground garages can weaken the effect.
Liu Yiquan believes that there are three main difficulties in the construction of high-precision maps. First, for the same spatial point, it may scan and generate data at different times and from different angles using different sensors. How to combine these “multi-source data” to ensure the consistency of map data is a difficult point; second, after all data acquisition is completed, in addition to data fusion, it is also necessary to automatically extract key features in the data; third, It is now possible to generate local high-precision maps, but it is still very difficult to generate super-large-scale high-precision maps.
High-intelligence mapping process demonstration (Source: high-sense map)
At present, the high-precision map of the high-intelligence map covers three scenarios: highways and urban expressways, urban roads, and enclosed areas.
Because the GPS signal is blocked, the underground garage is one of the most difficult scenes to make a high-resolution map. Traditional graphics companies generally choose to use more expensive equipment for data collection, millions of dollars, and DeepMap uses point cloud fusion based on low-cost sensors and ICP algorithm for map production, which has cost advantages.
Liu Yiquan told them that provides “relative accuracy” measurement methods and map generation methods, which can include GPS without “absolute precision” Information, Because absolute precision is actually an “observation average”, the error is inevitable and is constantly changing. At present, the map accuracy provided by Gaoshen Zhitu can reach 5cm, the positioning accuracy is less than 10cm, and the precision of the products on the market is generally about 20cm.
“Our maps and targeting are actually integrated, and we can understand our location service as a value-added service. You can just use our maps without our positioning, but if you use our Positioning, we must use our map. In fact, the combination of the two can make our technical ability to the best.” Liu Yuquan said.
Liu Yuquan told that Gaoshen Zhitu has no fixed data collection vehicle. It is a “light asset model”. Usually, the data is collected by the customer’s car or by himself. If the data size is large, for example, to collect a map of the entire Beijing City and keep it updated daily, they will cooperate with the third party, for example, looking for a taxi , bus or sanitation company, install the collection equipment on the roof of these cars.
Liu Yiquan said that the acquisition equipment they use is Lidar, camera, GPS and Inertial Navigation (IMU). The market price is about 30,000 US dollars, far lower than the equipment cost of most manufacturers.
In terms of data update, Gao Zhizhi is the first to propose “crowd mode”One of the “pictures” is the so-called “crowdsourcing”, that is, every auto-driving vehicle with sensor configuration can be the source of map data update. In order to ensure the real-time update of the high-end map of the car, high-intelligence map Deploy the entire artificial intelligence production process on the cloud and use the super-power on the cloud to handle the huge amount of data in the map production process.
Talking about whether autopilot companies can map their own drawings and do not need to outsource to map dealers. Liu Yiquan said that local areas may be possible, but if scaled up, it is very difficult. Moreover, many technical details are difficult to deal with. For example, when doing image semantic segmentation, it is necessary to have a good understanding of local traffic lanes, signage, signal lights and other traffic regulations and traffic signs. Moreover, looking at the entire autonomous driving industry, it is impossible for a company to complete all the links. The whole industry can develop if you work together and do what you are best at.
“Compared with all domestic competitors, Our strongest place isin Europe, USA, Japan, Australia, < /span>New Zealand and other parts of the world are drawing, this is one of our absolute advantages,” said Liu Yuquan, overseas. OEMs want to enter China, Chinese car companies want to go out to sea, if the map standards of the two places are the same, are provided by the same company, it is much more convenient for customers.
Talking about the current development bottleneck of high-precision maps, Liu Yuquan said that the surveying and mapping industry has strict policies, not all enterprises are allowed to enter, and there are many rules for applying for Grade A surveying and mapping qualifications. To a limited extent, the business of the graphics business has been limited, and the government’s legal requirements for “deflection encryption” have yet to be improved.
Liu Yiquan believes that high-precision maps are a matter of scale. The larger the scale, the lower the unit cost of data processing is , and after the autopilot is popular, car companies can use their own devices to go to And update the high-precision map , which saves the equipment cost.
Before joining the company, Liu Yiquan was the general manager of the Alibaba Cloud Global Technology Cooperation Division and the general manager of the Solution Division. He was also responsible for the Oracle China Enterprise Architect team.
————
I am a reporter Wang Yizhen, business communication can add WeChat catherineyijin, please note the name + company position + come.