Huajin Securities pointed out that the algorithm architecture of Huawei's intelligent driving has changed from manual annotation to autonomous decision-making, from having pictures to no pictures. Vehicle recognition from BEV to GOD, from manual annotation to autonomous decision-making.
BEV (Bird's Eye View) technology is an important way in the intelligent driving perception algorithm, which can be based on multi-sensor data fusion, usually using the image information collected by cameras and other sensors, after processing and analysis, the environment around the vehicle is presented from a bird's-eye view, so that the vehicle has an overall spatial perception of the surrounding roads, vehicles, obstacles, etc.
According to Huajin Securities, Huawei's ADS1.0 mainly uses the BEV+Transformer algorithm, which is a new neural network architecture that can directly convert between different sequences of 2D and 3D. The idea of the whole BEV+Transformer scheme is basically "input-extraction-transformation-fusion-timing-output", which can fuse and display the data collected by multiple sensors in the same coordinate to form a virtual vector space, in which all analysis and decision-making are carried out. However, BEV has certain limitations in the recognition of special-shaped obstacles and some irregular objects, and it is difficult to accurately identify and judge those objects that are not within the scope of its preset model and annotation. And in the development stage, the algorithm relies on the "whitelist mechanism", which requires manual labeling of the identified targets. However, the actual road conditions are extremely complex, and there is a "long tail effect" in obstacle marking. Objects that are not on the white list, such as rocks and trees, are not only numerous, but also diverse, which leads to accidents caused by intelligent driving systems that cannot recognize these objects.
Figure: Schematic diagram of BEV input and output (Source: Huajin Securities)
"Knowing the road": from there is a picture to no map, RCR is used to realize the normal driving of the vehicle
In the field of intelligent driving, accurate "road knowledge" has always been the core key. In the early days, intelligent driving relied on high-precision maps, which were like equipping vehicles with an accurate "navigation blueprint", and the vehicles relied on the detailed road information in the maps, such as lane locations and traffic sign locations, to achieve relatively stable driving. The HD map is like an experienced guide, providing various details for the vehicle during the driving process, helping the vehicle plan the route in advance and avoid potential dangers.
However, HD maps have many limitations. It is extremely expensive to produce and maintain, requires a lot of manpower and material investment, and needs to be constantly updated to match the dynamic changes of the road, such as road construction, temporary traffic control, etc., and there is often a lag in map updates. Moreover, HD maps have limited coverage and may not be able to provide effective data support in some remote areas or newly built roads.
To overcome the dilemma, Huawei has introduced Road Cognition & Reasoning (RCR) technology for intelligent driving to make the leap from mapped to non-mapped. RCR technology allows vehicles to rely on their own sensors, such as cameras, millimeter-wave radars, and lidars, to perceive and understand their surroundings in real time. Through advanced algorithms, vehicles can autonomously recognize road features, lane markings, traffic signs, and other vehicles and pedestrians, just as a human driver can make driving decisions based on their eyes observing the road.
In practice, RCR technology plays an important role when the vehicle is driving on a road that is not covered by a high-definition map. The sensor continuously collects information about the surrounding environment and transmits it to the vehicle's intelligent driving system, which quickly analyzes and processes these data to determine the vehicle's current position, driving direction and relative position relationship with surrounding objects, so as to plan a reasonable driving path and ensure the normal driving of the vehicle.
It is reported that Huawei's intelligent driving technology has been widely used. For example, in February 2024, Huawei's OAIT launched the "Navigation Control Assistant" (NCA) feature, which enables vehicles to perform urban driving assistance without relying on high-precision maps. At present, this feature has been supported in 338 cities in Chinese mainland.
Huawei predicts that by 2030, its intelligent driving algorithm will achieve a general object detection accuracy rate of more than 99%. In addition, Huawei will continue to optimize its intelligent driving technology and continuously improve the performance and user experience of the system through cloud computing and AI cluster training.
In short, Huawei's intelligent driving technology has changed from relying on high-precision maps to implementing mapless navigation, reflecting its strong technical strength and innovation capabilities in the field of autonomous driving. Through RCR technology, Huawei not only improves the vehicle's perception and decision-making capabilities, but also lays a solid foundation for the future development of intelligent driving.
Related:
Huawei Intelligent Driving Solution Market Analysis Report (1)
Huawei Intelligent Driving Solution Market Analysis Report (2)
Huawei Intelligent Driving Solution Market Analysis Report (3)
Huawei Intelligent Driving Solution Market Analysis Report (4)
