At present, most of the automotive autonomous driving solutions have achieved L2 and above autonomous driving through the BEV+Transformer architecture. Most car manufacturers on the market, such as Huawei, Xpeng, Ideal, HAOMO.AI, Zhuoyu Technology, Jiyue and other companies have achieved high-speed NOA (an automatic driving assistance system that can make drivers drive safer and more efficiently on highways) through BEV+Transformer+OCC. Urban NOA (Driver Assistance System developed for urban traffic environments. As well as support for automated parking. The BEV+Transformer architecture is an advanced perception technology for autonomous driving that enables more accurate perception of the vehicle's surroundings by converting 2D images captured by multiple cameras into 3D representations of bird's-eye views. The core advantage of this architecture is that it can process and fuse spatial information and time series information from different perspectives to provide richer and more accurate perception data for autonomous driving systems.
EO Intelligence believes that by 2027, the field of autonomous driving is expected to realize the whole process of perception, planning, and control, and realize autonomous driving based on learning-based.
Figure: Iteration of the architecture of autonomous driving intelligent algorithm (Source: EO Intelligence)
According to the analysis of the iteration of the autonomous driving algorithm architecture of EO Intelligence, the autonomous driving in 2016 relied on the manually designed feature extraction method, which had poor prediction ability, mainly relied on pre-designed features and experience summary, and only supported L2 level of autonomous driving and ADAS assisted driving functions. By 2024, vehicles can automatically extract features through self-attention mechanisms combined with the diffusion process, and provide an understanding of the environment through simulation and prediction of physical rules in some scenarios. Have a strong predictive ability to make predictions based on current perception, possible future occupancy states, and understanding of parts of the environment. It supports L3/L4 autonomous driving, high-speed NOA, urban NOA, automatic parking scenarios and other scenarios. According to the prediction of EO Intelligence, it is expected that by 2027, self-driving cars will automatically extract features through self-attention mechanisms combined with diffusion processes, and improve their understanding and reasoning capabilities of scenarios through simulation and prediction of physical rules combined with visual and linguistic transformation. In addition, there is a strong predictive ability to make predictions based on current perception, possible future occupancy states, and a deep understanding of the environment. It supports L4/L5 autonomous driving, supports the integration of all scenarios of autonomous driving, and gradually realizes zero takeover.
Figure: Integrated development of mushroom car intermodal road in Yunnan (Source: EO Intelligence)
Information Fusion Vehicle-road-cloud integrated development
According to the EO Intelligence, the higher the level of autonomous driving intelligence, the more information about perception, measurement and fusion, the more computing power on the decision-making side, and the stronger control ability on the control side. Compared with single-vehicle intelligence, the level of autonomous driving of vehicle-road-cloud integration is significantly improved. Single-vehicle intelligence refers to obtaining external information through the sensors of the vehicle itself and processing it through algorithms, so as to realize automatic control of vehicle driving, which has high requirements for vehicle-end sensors and vehicle-end computing platforms. The vehicle-road-cloud integration is to obtain external information from vehicle-end perception, road-end perception + cloud perception, and achieve a higher level of autonomous driving through multi-terminal decision-making and multi-terminal control after information fusion.
According to the analysis of EO Intelligence, AI technology runs through the whole system of MOGO connection products, based on the group intelligence system architecture, complete independent research and development of the vehicle-road-cloud integration solution, open up the three ends of the vehicle-road-cloud, and realize the closed loop of data. At present, MOGO has many years of experience in the deep cultivation of vehicle-road cloud, and has achieved rapid implementation in multiple places and scenarios, covering urban open roads, highways, parks, scenic spots, ports, airports, etc.
Relateds:
