Qingzhou Zhihang Hou Cong: the new trends of the "four modernizations" in the evolution of intelligent driving

Editor's note: On December 16, the AutoHeart "2022 Smart Car Technology and Business Innovation Forum" was successfully held. With the theme of "New Technology and New Business", it focused on car intelligence and gathered academic insights from the smart car industry and the autonomous driving industry chain. Leaders, industry figures, entrepreneurs, and investment tycoons participated in the meeting and shared their views on the era of smart cars and the technological and business innovation opportunities in automotive technology. Through layer-by-layer analysis and trend judgment, they explored the future of China's smart car innovation era.

At the conference, Hou Cong, co-founder & CTO of Qingzhou Zhihang, delivered a speech titled "New Trends in the "Four Modernizations" of Intelligent Driving Evolution" and shared Qingzhou Zhihang's thinking on building urban NOA and its recent achievements.

The following is a transcript of the speech, compiled by AutoHeart:

Hello everyone, I am Hou Cong, the co-founder and CTO of Qingzhou Zhihang. I am very happy that AutoHeart has provided us with a platform so that we can learn from each other and exchange experiences with everyone in the automotive industry and intelligent driving industry.

The title of my speech today is "New Trends in the "Four Modernizations" of Intelligent Driving Evolution". I will talk about how Qingzhou Zhihang builds NOA, a city with Chinese characteristics.

Qingzhou Zhihang is now more than just an L4 autonomous driving company. We released a new dual-engine strategy in May this year.

"Double Engine" refers to the power engine and the innovation engine. The power engine refers to our autonomous driving technology capabilities and R&D system.

On top of the power engine, we continue to make breakthroughs and innovations to consolidate our technological depth and support long-term development.

The innovation engine is based on commercialization of our autonomous driving capabilities.

At present, we mainly focus on high-end assisted driving in front-end mass production, thereby broadening the width of our scenarios, extensively accumulating data, and feeding back our technological improvements.

We hope to rely on the power engine technology base to promote the commercial implementation and large-scale verification of innovative engines, and achieve our ultimate goal through this route - bringing autonomous driving into reality. The focus today is to talk to you about how our innovation engine brings urban NOA into reality.

When it comes to assisted driving, we believe that assisted driving systems that can bring value to users can be divided into three levels: "usable", "easy to use" and "loved to use". This is based on the stage of technological development and the scope of application of assisted driving. to distinguish. in:

• 
  

• "Usable" is mainly the basic high-speed NOA function, which is an auxiliary driving that has relatively high requirements on road conditions; "easy to use" means that it can achieve a better high-speed NOA experience and can be enabled in most high-speed scenarios, and in some cases It can also be activated on urban sections with better conditions. "Love to use" refers to the realization of urban NOA within a large ODD range, point-to-point automatic driving, which can achieve a good driving experience in a wide range of scenarios, allowing assisted driving to help driving as easy as hailing a taxi, and forming a user-friendly environment. rely.

We believe that the scope of application of urban NOA should include urban roads, highways and expressways, and parking.

In these scenarios, the ability to achieve point-to-point, good assisted driving functions can be called urban NOA. This is also its value. It can help drive more efficiently and easily, and achieve a state of human-machine co-driving.

The reason why Qingzhou attaches great importance to and insists on doing a good job in urban NOA is because we believe that whether it is technical capabilities or application scenarios, it is the ceiling of assisted driving and the most pioneering product function of automobile intelligence.

When our urban NOA can be used in more than 90% of the scenes in the city, we can make breakthroughs in this assisted driving capability towards driverless driving. Through urban NOA, we can continuously accumulate data in a wide range of scenarios, and we can also Continue to improve technology and move closer to driverless driving.

Therefore, urban NOA is the entry threshold for autonomous driving and the only way to achieve the goal of autonomous driving.

Some market analysis data also show that urban NOA has now become the general trend for high-end intelligence in automobiles.

Data shows that by 2025, the market penetration rate of L2 and above assisted driving will rise to nearly 40%, and consumers are increasingly paying attention to vehicle intelligence, reducing fatigue, freeing hands, increasing safety and other functions, indicating that high The value brought by high-level assisted driving will also receive more and more attention and gradually become what everyone expects.

So why can Qingzhou be a good urban NOA?

We believe that Qingzhou has the three elements of time, location, and people.

• 
  

• Timing is the industry trend and consensus;

• 
  

• The geographical advantage means that in the booming automobile and technology industry chain, many excellent partners have emerged. On our road to mass production and commercialization, we have gradually formed a very good alliance with many hardware, software, OEM and other companies. Ecological partner circle;

• 
  

• Renhe means that Qingzhou, as an L4 company, has sufficient autonomous driving R&D experience and technology accumulation, as well as a professional R&D team. Therefore, Qingzhou is very confident that it can do a good job in urban NOA.

In terms of assisted driving, we mainly divide it into three levels: high, medium and low according to hardware cost and sensor configuration:

• 
  

• The low-end ones are mainly based on monocular cameras and are purely visual solutions; the mid-range ones are also purely visual, but they will have more equipment; the high-end ones will add a lidar. We believe that if we want to implement urban NOA functions in China, we need a high-end configuration.

We have also been thinking about Tesla's FSD and what configuration it should have in China. We think it is a mid-range solution in terms of hardware, and it is a purely visual solution. However, FSD can achieve quasi-high-end performance in North America because the scenes abroad are relatively simple, the roads are relatively regular, and the rules are relatively complete.

However, China's road conditions are relatively complex. Not only are the roads not regular enough, but the driving is also lack of standardization. Pure visual detection has certain flaws in identifying obstacles. These situations can be dealt with well if used with lidar.

Whether a house is built well or not depends first on whether the perceived foundation is solid.

Let’s first introduce the algorithm of Qingzhou’s perception. Qingzhou sensing is mainly based on the concept of hyper-converged sensing, specifically multi-sensor timing interleaved fusion. We know that fusion is divided into three stages: pre, middle and post, which mainly refers to the fusion of data sets, feature sets and target sets.

Qingzhou's super fusion realizes all the front, middle and rear fusion, and also adds timing fusion.

Taking medium fusion as an example, it fuses the data from lidar and camera in the BEV space on the feature layer, and then supplements it with a fusion in time series to generate more accurate perception results, and can combine the objects' Speed ​​and direction are estimated.

Under the concept of hyper-convergence, in order to accelerate the mass production of urban NOA, Qingzhou deployed a large model of temporal multi-modal feature fusion on the mass production platform of Horizon Journey 5 for the first time in the industry.

Here I will introduce Qingzhou’s large perception model called OmniNet, which is a full-task large model used in the front and middle fusion stages to achieve data and feature fusion of vision, millimeter-wave radar, and lidar.

Omni is the abbreviation of the English word "omnipotent". This model will serve as the main model, supporting almost all core perception functions, enabling efficient multi-task unified computing, and can output the results of multiple tasks at the same time.

For mass production deployment of OmniNet, it has three core advantages:

The first is more accurate perception. Through multi-sensor, multi-level, and timing fusion, OmniNet can output rich and accurate environmental perception results, and the multi-task output can also complement each other.

In terms of fusion strategy, OmniNet fuses image information from various perspectives in the BEV space. It can accurately and stably identify ultra-long vehicles, special-shaped vehicles, or cross-camera truncation objects, etc., and assists the model in subsequent joint sequential multi-sensor fusion.

Through the BEV space, image field information can also be projected on the 3D information of lidar, achieving a more accurate, semantically rich, and clearer perception effect.

The second advantage is that the car end is more suitable. OmniNet allows originally independent computing tasks to perform efficient and multi-task unified computing by sharing the backbone network, which can save about 2/3 of the computing power and better meet the application needs of automotive-grade chips.

In addition, the sensors here can also be configured, such as lidar, which can use one or multiple sensors. Even if the purely visual solution of lidar is removed, such a network structure can still be reused.

Finally, iteration is more efficient. OmniNet is a data-driven development model that shortens the algorithm iteration cycle and reduces model maintenance costs through closed-loop data methods such as self-supervision and semi-supervision. At the same time, the effect is more reliable, which is conducive to handling various problems in urban operations. Such a long tail problem.