Learn more about finance |BT Financial Data Intermedia
Whether a company's technology is strong or not depends on whether it can be launched and can be searched by users.
Huawei, which is a leader in the field of intelligent driving, on February 2, 2024, has no map for intelligent driving 2The OTA upgrade of 0 function has completely made the owners of Wenjie, AVATAR, Jihu, and Zhijie boiling.
In this list of cities that have officially opened up for intelligent driving, people are surprised to find that the number of available locations covering towns and villages exceeds 720,000, covering 99% of the country's road sections, truly realizing intelligent driving that can be used in the country.
What's more interesting is that Xpeng, which was previously considered to have the most intelligent driving in the city, now ranks second, and the number of cities that can be used is only more than 200.
The comparison of the overall data, as well as the actual measurements of a large number of car owners, make all those who care about China's auto industry feel speechless, and it is indeed far ahead.
People especially want to know how Huawei, which has not been completely completed by Xiaopeng and other companies for many years, has been put into use in the whole country, and Huawei, which is the latecomer, has done it?
In the frontier field of automotive technology, intelligent driving technology has become a research hotspot. This technology uses cutting-edge sensors and computer technology to simulate the decisions and actions of a human driver, giving the vehicle the ability to "think". Imagine an intelligent driving system like a mechanical driver, constantly monitoring and scanning the surrounding environment through sensors and radars installed around the vehicle, realizing the dream of autonomous driving.
In the past intelligent driving implementation schemes, high-precision maps played a crucial role.
Compared with ordinary navigation maps, high-precision maps have higher accuracy and more information. While normal maps can only achieve road-level accuracy, high-definition maps can achieve decimeter or even centimeter-level accuracy, including accurate lane markings, traffic lights, signs, and other details. On the basis of the high-definition map, a dynamic map will be superimposed, and the surrounding environment information will be collected and annotated in real time through sensors.
However, HD maps, while important, come with a number of challenges.
First of all, it takes a lot of money and time to make HD maps. According to industry insiders, it costs billions of yuan or more to collect high-precision map data across the country. In addition, HD maps are updated relatively slowly, making it difficult to keep up with reality. Due to road changes, new construction, and other factors, HD maps are prone to deviation or obsolescence.
According to statistics, the mileage of China's expressways is about 300,000 kilometers, and the mainstream map dealers have basically completed the high-precision map collection of expressways and urban expressways, which is also the basis for the high-speed pilot assisted driving products of the above-mentioned technical route manufacturers.
But there are two significant differences between urban roads and highways. One is that the road mileage is longer, according to statistics, close to 10 million kilometers, which is more than 30 times the mileage of the expressway, while the mainstream map merchants have only completed about 200,000 kilometers of data coverage (concentrated in Beijing, Shanghai, Guangzhou and Shenzhen). The other is the high degree of road variability, such as irregular maintenance of urban roads, dense large vehicles, irregular parking of roadside vehicles, and frequent encounters with open car doors.
Therefore, if you want to realize the urban navigation assistance driving function based on high-precision maps, the collection ability of map merchants and the freshness of high-precision maps are very high, and the freshness should reach the level of time or day change. However, at present, the high-definition map of the city that can be provided by the map provider can only be updated on a monthly or quarterly basis. Based on this degree of high-precision map, products can also be made, but they cannot meet the strict quality control requirements of OEMs.
In response to these problems, some companies have begun to explore mapless intelligent driving solutions. The so-called no-map means that instead of relying on high-definition maps, the car can perceive the surrounding environment in real time through sensors such as cameras and radars, and think and judge according to the thinking mode of the human driver. This approach requires extremely high algorithms, as the vehicle needs to process a large amount of environmental information in real time and make accurate decisions.
In the mapless intelligent driving solution, environmental perception, positioning and control path are the three core technologies. Environmental perception mainly relies on sensors such as cameras and radars to detect road information such as lane markings, road teeth, and stop lines in real time. Positioning technology fuses a variety of sensor information to achieve accurate positioning of the vehicle in a real-time map. The regulatory route is based on information such as navigation, environment, and vehicle movement position to plan the best driving path.
The realization of the mapless intelligent driving solution is inseparable from advanced algorithms and powerful computing power. Vehicles need to process massive amounts of environmental information in real time and make decisions in milliseconds. This puts forward higher requirements for the optimization of algorithms and the improvement of computing power.
The difference between mapless and mapped intelligent driving is that the input of the original high-precision map is replaced by the input of the navigation map and the understanding of navigation information by real-time perception. Due to the current mapless scheme, whether it is for vision or BEV+Transformer detection scheme, the accuracy requirements for lane line perception are more accurate, and the system can further perceive complex road structures by learning a large number of road and intersection features, and finally make judgments.
That's what Xpeng and other car companies do. But Huawei's gameplay is different.
Recently, Huawei's patent layout in the field of intelligent driving has also attracted widespread attention. According to the announcement of the State Intellectual Property Office, Huawei Technologies has applied for a patent called "an intelligent driving method and a vehicle applying the method". The patent describes a dynamic decision-making method based on perceptual information, which can effectively deal with the problem of inconsistency between the decision-making result and the actual action of the game target, so as to enhance the generalization ability of game decision-making and improve the safety and driving comfort of intelligent driving.
At the forefront of intelligent driving technology, Huawei's pictureless NCA intelligent driving system is leading a technological revolution.
This set is based on ADS2The 0 system, with perception as the core, completely gets rid of the shackles of high-precision maps, integrates the innovative BEV network and the industry's first GOD network, and brings new possibilities for intelligent driving.
The BEV network, figuratively known as the "God's Perspective", uses a whitelist to identify obstacles and provides a full range of perception capabilities for vehicles. The GOD network is closer to people's visual habits, it does not rely on whitelists, but through the planning of passable areas, so that vehicles can be upgraded from "visible" to "understandable". The design idea of this network not only improves the accuracy of recognition, but also greatly enhances the vehicle's adaptability to the environment.
Huawei's intelligent driving mainly relies on two algorithms:
"Understandable" GOD (General Obstacle Detection) 2As the "eye" of Huawei's intelligent driving system, the 0 network has powerful recognition capabilities.
It can not only identify special-shaped objects outside the general obstacle whitelist, but also finely identify the types of obstacles, such as distinguishing ambulances, police cars, etc. This innovative recognition method breaks the dependence of traditional intelligent driving systems on fixed obstacle identification, enabling Huawei's intelligent driving system to respond more flexibly to various complex road conditions. At the same time, God 2The recognition rate of the 0 network is as high as 999%, which provides a very high safety guarantee for intelligent driving.
RCR (Road Cognition & Reasoning) 20。This network is designed to match navigation maps with the real world to provide accurate path planning for intelligent driving.
To solve the problem of identifying special-shaped objects and unlabeled objects, Huawei has adopted a similar technical approach as Tesla. However, Huawei's deep accumulation of LiDAR technology enables it to better integrate perception sensor cameras and LiDAR into a spatiotemporal network. This spatiotemporal network contains both spatial and temporal information in its structure and analysis, enabling the modeling and analysis of complex systems that evolve over time and space. In this way, Huawei not only improves the accuracy of identifying special-shaped objects and unlabeled objects, but also reduces the requirements and shortcomings for the computing power of vision algorithms.
Another rcr 20 network is better
rcr 2.0 network does not require the existence of intelligent driving maps at all. This innovative technological breakthrough breaks the industry's dependence on special intelligent driving maps.
In contrast, many intelligent driving solutions that are shouting to go to high-precision maps are still needed, such as the HQ map pushed by AutoNavi, HD Air pushed by Tencent, and the newly promoted functional map of NavInfo.
Huawei, on the other hand, uses the SD map information of human driving to achieve a perfect match between the navigation map and the real world. This innovative technical route not only simplifies the complexity of intelligent driving systems, but also reduces their costs and application thresholds.
The high-precision map provides a reference line for high-precision driving of the car, and the intelligent car travels according to the reference line with high-precision positioning, regardless of the turn or intersection according to the line, which you can understand as an electronic track similar to a train. Therefore, when there is no high-precision map, the intelligent driving of the vehicle will need to solve two difficulties by itself:
How do I know if a vehicle is in a lane? Generally, it is visual lane line recognition technology; And how to determine the relationship between lanes at intersections to ensure that you turn or walk straight in the right lane?
Huawei should use Tesla's similar vision algorithm technology to determine the vehicle driving in the lane, and based on the route information provided by the SD navigation map used by humans and the visual perception information of the intersection, the RCR deduces the relationship between the intersection lanes, so as to achieve pilot assistance.
As a result, the composition of these algorithms has what Huawei calls both the GOD and the RCR algorithm that can "understand the road". In general, it is similar to Tesla's route, but Huawei has played lidar with yo-yo, adding more secure data matching to ensure the safety of intelligent driving.
Huawei's intelligent driving solution can quickly achieve safe driving and use in Kaecheng across the country, which is inseparable from the introduction of the Pangu model into intelligent driving.
The Pangu model's support for Huawei's intelligent driving solution is actually in data processing. In the past, the data processing of autonomous driving had to go through the method of algorithm + manual annotation, which was time-consuming and laborious, and the timeliness was not easy to guarantee, so that users could not truly ensure safety.
The main role of the Pangu model in Huawei's intelligent driving solution is to improve the data closed loop.
The core purpose of data closed-loop is to continuously extract the first-class data from the original data, send it to the algorithm for training, and finally continuously improve the algorithm to solve various long-tail scenarios. The method is to extract data from the vehicle end for initial sorting, then carry out scenario-based management, and then carry out data labeling, training, and finally return to the collection needs. Among them, every process has an opportunity point where big data can exert force.
By building a digital twin and generating samples of complex scenes, the Pangu automobile model can accelerate the learning and response ability of the autonomous driving system to complex scenes. This technological innovation can shorten the learning and training period of autonomous driving from more than two weeks to two days, greatly improving the iteration speed and application efficiency of autonomous driving technology.
In layman's terms, it is a variety of information collected on the spot, and in the past, autonomous driving was processed by the idea of algorithm and manual setting, which had a strong sense of mechanics and could not accurately respond to various emergencies on the road. Now, after Huawei introduces the Pangu large model, the data collected by any sensor is processed by the device-side + cloud large model, which can quickly give decision-making opinions and finally form autonomous driving decisions.
In addition, Huawei's Pangu model has a particularly strong learning ability, and after passing through the same road section, it will not only accept road information, but also record user habits, which will make his operation quickly close to the driver's driving habits, and finally produce an autonomous driving mode with thousands of people.
This is also the reason why Huawei's smart driving ads now show that "the same road section is better to drive for the second time".
Also, look at a set of data:
The OTA covers 4 municipalities, 43 regions, 290 prefecture-level cities, 1,636 counties, 374 county-level cities, 14,677 townships, 19,531 towns, and 691,510 administrative villages, with a total of 728065 cities.
Because the data released by the Passenger Car Association shows that in 2023, the penetration rate of intelligent driving of passenger cars will grow inversely with **, and the penetration rate of intelligent driving of L2 and above in the passenger car market will reach 424%, 70% is expected in 2025, and it will be widely used in mainstream models of 100,000 to 200,000 yuan.
It is not difficult to see that consumers are not as reluctant to intelligent driving as in the past, and other car companies are also increasing investment in the field of intelligent driving. It is foreseeable that this year is undoubtedly the most competitive year for intelligent driving.
In this context, Huawei is the first in the country to achieve full coverage of 99% of roads without maps, and it can only be said that "far ahead" is not a slogan.
This in itself is outrageous, but such an outrageous thing is done by Huawei, which makes people feel that "technology research and development is the primary productive force".
Kunpeng Project
Author | rickzhang