This paper argues that intelligent automotive technologies, exemplified by autonomous driving systems, will significantly impact the transformation of the automotive industry ecosystem. First, it analyzes the development path and current status of autonomous driving technology, as well as the application areas of Artificial Intelligence (AI), highlighting the challenges AI faces in autonomous driving applications. Then, it proposes an AI-based vehicle-cloud collaborative autonomous driving system architecture, analyzing the hardware and software architecture of AI-based intelligent driving terminals and the cloud-based architecture of autonomous driving based on big data. Finally, it discusses the main applications of artificial intelligence in autonomous driving systems, considering the integration and application of AI between the vehicle and cloud, information and data interaction methods, and vehicle-cloud collaborative technologies.
Since the beginning of the 21st century, with the accelerated integration of new-generation information and communication, new energy, and new materials technologies with the automotive industry, and with emerging technology companies in information and communication, and the internet, fully entering the automotive sector, the global automotive industry is facing significant changes.
1) Profound transformation in product form and production methods. Automobile products are rapidly developing towards new energy, lightweighting, intelligence, and connectivity. Cars are transforming from transportation tools into large mobile intelligent terminals, energy storage units, and digital spaces. Automobile production methods are evolving towards a fully interconnected and collaborative intelligent manufacturing system, with personalized customization becoming a trend.
2) Emerging demands and business models are rapidly emerging. User experience has become a crucial factor influencing car consumption. Consumer demand is becoming increasingly diversified, with shared mobility and personalized services emerging as key trends.
3) Profound adjustments in the industrial landscape and ecosystem. Developed automotive countries are accelerating industrial innovation and integration, leading to rapid changes in the global automotive industry landscape. Emerging technology companies, including those in the internet sector, are entering the automotive industry in large numbers, reshaping the global automotive ecosystem.
Intelligent automotive technology integrates modern sensing technology, information and communication technology, automatic control technology, and artificial intelligence, and has enormous potential in reducing traffic accidents, alleviating traffic congestion, reducing energy consumption, and protecting the environment. Therefore, countries around the world are actively developing roadmaps for autonomous driving technology and promoting the development of autonomous vehicles. Various science and technology plans, such as the US's "Industrial Internet," Germany's "Industry 4.0," and Japan's "Robot Revolution," all list intelligent automotive technology as a key breakthrough for the development of the automotive industry. my country's "Made in China 2025" (released in 2015) and "Medium- and Long-Term Development Plan for the Automotive Industry" (released in 2017) explicitly propose a "smart + connected" development path for autonomous vehicle technology.
Autonomous driving systems are considered the ultimate goal of automotive intelligence development. They are of great strategic significance for improving traffic safety, achieving energy conservation and emission reduction, eliminating congestion, improving social efficiency, driving the coordinated development of automobiles, electronics, communications, services, and social management, and promoting the transformation and upgrading of the automotive industry. Autonomous driving technology has become a hot spot of competition for many companies.
On the other hand, given the successful applications of deep learning (DL) in machine vision (MV), natural language processing (NLP), and other fields, and my country's AI development strategy, researching the deep application of AI in autonomous driving systems is of great practical significance. This paper reviews the development trends of autonomous driving technology, analyzes the key technologies of artificial intelligence in the development of intelligent and connected vehicles, and proposes an AI-based vehicle-cloud collaborative autonomous driving system.
1. Analysis of the Development Trends of Autonomous Driving Technology and the Current Status of AI Applications
1.1 Current Status of Autonomous Driving Technology Development
An autonomous vehicle is a complex cyber-physical system that is mechatronic, highly integrated with software and hardware, and ultimately aims to replace human operation. It mainly consists of perception, decision-making, and execution subsystems. Autonomous driving technology involves key technologies such as environmental perception, decision-making and planning, control execution, and V2X communication. Its structure is shown in Figure 1.
Figure 1. Schematic diagram of the layered structure of an autonomous vehicle
Environmental perception technology uses onboard sensing devices (such as GPS/INS systems, millimeter-wave radar/cameras) and 5G networks to acquire traffic environment information and vehicle status information (position, attitude) of the vehicle. It also unifies the output information of multiple sensors under the vehicle coordinate system to establish meta-information with time stamp data association and fusion, which serves the decision-making and planning of autonomous driving.
The decision-making and planning technology uses information output from the environmental perception subsystem to achieve functions such as route finding, traffic prediction, behavior decision-making, action planning, and feedback control signal output.
Control execution technology uses drive-by-wire actuators to execute feedback control output commands in order to control steering, throttle, and braking.
V2X technology provides real-time and reliable communication services for information interaction between vehicles and the outside world, serving environmental perception and decision-making planning.
1.1.1 Development Roadmap for Autonomous Driving
Currently, there are two paths to realizing autonomous driving technology: a gradual development path mainly led by traditional automakers, and a disruptive development path mainly led by research institutions and IT companies.
1) A gradual development path involves progressively improving the level of vehicle intelligence, progressing in stages towards assisted driving, partial autonomous driving, highly automated driving, and fully automated driving. In the assisted driving stage, vehicle control is primarily driver-centric, with the driver retaining ultimate driving control. The system assists the driver, reducing their workload. Currently, assisted driving technologies mass-produced in passenger vehicles include lateral stability control and electric power steering control. Some high-end vehicles also feature automatic parking, adaptive cruise control, and lane departure warning systems. In the partially automated driving stage, the vehicle's intelligence level further improves, possessing a certain degree of autonomous decision-making ability and the capacity for short-term management under specific conditions. In the highly automated and fully automated driving stages, the vehicle exhibits a high degree of autonomy, capable of autonomous decision-making, planning, and control, enabling management in complex conditions (such as highways and urban environments) and even fully driverless operation.
2) The disruptive development path skips the gradual development of automotive intelligence, directly achieving highly/fully autonomous driving. This approach is highly challenging, but its research results have been effectively applied to various stages of the incremental development path. The United States is the earliest and most technologically advanced country in this field. Since the 1980s, the U.S. Defense Advanced Research Projects Agency (DARPA) has funded U.S. companies, research institutions, and universities through projects such as ALV, DEMO-II, and DEMO-III to apply disruptive autonomous driving technology in the military field. Google is currently the most successful company in this area. It began researching autonomous driving technology in 2009, conducted urban road tests of autonomous vehicles in 2010, and obtained authorization for autonomous vehicles in 2011. Its autonomous vehicles are currently considered compliant with federal law by U.S. vehicle safety regulators. Germany was also among the first countries to begin research in this field. As early as the 1980s, the Federal Armed Forces University of Munich collaborated with Mercedes-Benz to develop autonomous vehicles. Its flagship model, the Mercedes-Benz S500, completed long-distance autonomous driving tests on urban and intercity roads in 2013.
The gradual and disruptive development paths reflect the divergence between traditional automakers and internet IT companies regarding the industrialization of intelligent driving vehicles. Internet companies attempt to introduce cutting-edge IT technologies into the automotive field to provide consumers with a better driving experience, using a top-down technology diffusion approach to vertically derive lower-level intelligent driving technologies. Automakers, on the other hand, believe that drivers' needs for safety may far exceed the expectations of IT thinking, and therefore adopt a gradual approach to promoting intelligent driving technology.
Regardless of the technological approach, research on vehicle intelligent safety assistance functions is already quite mature, objectively laying the foundation for intelligent driving. Both OEMs and IT companies face technical, cost, and regulatory challenges in the industrialization of intelligent driving. However, as long as the market continues to demand these technologies, it will drive the evolution of automobiles towards full driverlessness.
1.1.2 Current Status of Autonomous Driving Technology Development in my country
my country's research in the field of autonomous driving began in the 1980s. In 1980, the "Remotely Driven Nuclear and Chemical Reconnaissance Vehicle" was approved as a national project. In 1989, my country's first intelligent car was successfully developed at the National University of Defense Technology. In 1992, the National University of Defense Technology, Beijing Institute of Technology and other universities successfully developed my country's first truly autonomous test vehicle (ATB-1).
Entering the 21st century, the National 863 Program began to provide more support for research on autonomous driving technology. In 2000, the National University of Defense Technology announced the successful testing of its fourth-generation autonomous vehicle. In 2003, the National University of Defense Technology and FAW jointly developed an autonomous vehicle—the Hongqi CA7460—which could automatically change lanes based on road conditions ahead. In 2006, the new generation Hongqi HQ3 autonomous sedan was successfully developed. In 2005, my country's first urban autonomous vehicle was successfully developed by Shanghai Jiao Tong University. In 2011, the HQ3, developed by the National University of Defense Technology and FAW, completed its first fully driverless test on a highway from Changsha to Wuhan, achieving an average speed of 87 km/h and a distance of 286 km. In November 2012, the autonomous vehicle developed by the Military Transportation Academy completed highway testing, becoming the first officially certified driverless vehicle in my country, and winning the championship in the 2015 and 2016 China Intelligent Vehicle Future Challenge.
In December 2015, Baidu, an IT company, completed autonomous driving tests on open highways in Beijing, marking the transition of autonomous driving technology from research to product. In September 2016, Baidu announced that it had obtained the 15th global license for driverless car testing issued by the California government. On April 17, 2017, Baidu showcased a demonstration vehicle with enhanced highway assist features developed in cooperation with Bosch.
In April 2017, my country included intelligent connected vehicles based on autonomous driving technology in the "Medium and Long-Term Development Plan for the Automotive Industry," making it another strategic goal for the transformation and development of my country's automotive industry. However, the overall level of my country's autonomous driving technology still lags behind advanced international levels. Key technologies (perception fusion, path planning, control and decision-making technologies, etc.) are still in the development stage. Limitations in the development of these key technologies restrict the autonomous driving capabilities of autonomous driving systems in different environments, leading to sometimes significant discrepancies in the behavior of these systems.
1.2 Current Status and Challenges of AI Applications in Autonomous Driving
1.2.1 Introduction to Artificial Intelligence Technology
AI is a science that studies the theories, methods, and technologies for simulating, extending, and expanding human intelligence. It originated in the 1950s and has now developed into six major fields: computer vision, natural language understanding and communication, cognition and reasoning, robotics, game theory and ethics, and machine learning, showing a trend of mutual penetration among these fields.
Machine learning studies how to automatically learn the data structure and inherent patterns of input data samples under the guidance of algorithms, and gain new experience and knowledge to intelligently identify new samples and even predict the future. Typical machine learning algorithms include linear regression, K-means, K-nearest neighbors, principal component analysis, support vector machines, decision trees, and artificial neural networks.
Deep learning models, developed based on artificial neural networks, are among the most effective machine learning algorithms currently available, becoming a hot topic in artificial intelligence research and applications. Deep learning models incorporate multiple hidden layers into artificial neural networks and were first proposed in 2006 by Geoffrey Hinton and Ruslan Salakhutdinov. Due to their outstanding performance in the 2012 ImageNet competition (the most influential international competition in computer vision), deep learning models have received significant attention from all sectors of society and have made research progress in multiple fields, resulting in a number of successful commercial applications, such as Google Translate, Apple's Siri, Microsoft's Cortana personal voice assistant, Ant Financial's facial recognition technology, and Google's AlphaGo.
1.2.2 Application of Artificial Intelligence in Autonomous Driving Technology
AI has a wide range of applications in autonomous driving technology, with deep learning and reinforcement learning achieving good research results in the field.
1) Environmental perception field
Perception processing is a typical application scenario for AI in autonomous driving. For example, pedestrian detection technology based on HOG features typically uses support vector machine algorithms to detect pedestrians after extracting HOG features from images; in vehicle detection technology based on LiDAR and cameras, clustering processing of LiDAR data is required; linear regression algorithms, support vector machine algorithms, and artificial neural network algorithms are also commonly used for lane line and traffic sign detection.
Figure 2. Unstructured road detection framework based on machine learning
The framework shown in Figure 2 applies machine learning to the detection of unstructured roads such as rural highways and dirt roads. Due to the complex driving environment, existing perception technologies are insufficient to meet the needs of autonomous driving in terms of detection and recognition accuracy. Deep learning-based image processing has become a crucial support for the visual perception of autonomous driving. In the perception fusion stage, commonly used AI methods include Bayesian estimation, statistical decision theory, evidence theory, fuzzy reasoning, neural networks, and production rules.
2) Decision-making and planning field
Decision-making and planning is another important application scenario for AI in autonomous driving, with state machines, decision trees, Bayesian networks, and other AI methods already widely used. Deep learning and reinforcement learning, which have emerged in recent years, can achieve decision-making in complex situations through extensive learning and can perform online learning optimization. Due to the significant computational resources required, they are currently popular technologies in the computer science and internet fields for research on autonomous driving planning and decision-making.
3) Control execution domain
Traditional control methods include PID control, sliding mode control, fuzzy control, and model predictive control. Intelligent control methods mainly include model-based control, neural network control, and deep learning methods.
For example, Li Keqiang and others from Tsinghua University studied single-vehicle multi-objective coordinated adaptive cruise control technology, which comprehensively improves driving safety, improves vehicle fuel economy, and reduces driver fatigue while achieving three major functions: automatic following, low fuel consumption, and conformity to driver characteristics. They also proposed a cooperative multi-vehicle platoon control scheme based on a multi-agent system to achieve the goals of reducing fuel consumption, improving traffic efficiency, and enhancing driving safety.
1.2.3 Challenges Facing AI Applications in the Autonomous Driving Field
Currently, contemporary AI technologies, represented by deep learning, have been successfully applied in fields such as machine vision (MV) and natural language processing (NLP), and have been introduced into research on environmental perception, decision-making, planning, and control execution in autonomous driving technology, achieving good results.
Due to the complex driving environment, some AI technologies that heavily rely on data, computing resources, and algorithms are still unable to meet the real-time requirements of autonomous driving in areas such as perception, decision-making, and execution. This presents challenges for some autonomous driving system prototypes that rely on these technologies as their core support.
1) Real-time reliability requirements pose challenges to the system's computational speed and reliability. Autonomous driving systems require that the responses of the perception, decision-making, and execution subsystems be reliable in real time, thus necessitating the system to provide high-speed and reliable computing capabilities.
2) The industrialization demand for component miniaturization poses a challenge to the current large hardware size of the system. Most current autonomous driving system prototypes are computer systems or industrial control computer systems, which do not meet the requirements of automotive-grade components.
3) Personalized adaptation cannot be met. Current emerging deep learning algorithms have poor adaptability to variations in application environments, and there is an adaptation problem of model retraining for different vehicle models and different scenarios, which existing autonomous driving system prototypes cannot meet.
4) The requirements for autonomous learning and maintenance cannot be met. Deep learning exhibits the characteristic that the larger the learning set, the better the effect. Therefore, autonomous driving systems need to have continuous autonomous learning capabilities, which existing autonomous driving prototypes cannot meet. Faced with problems such as aging and wear, the calibration parameters of components at the time of manufacture are no longer in the optimal state. Autonomous driving systems need to perform intelligent tuning (self-calibration), diagnosis, and maintenance based on vehicle driving data and performance evaluation. Existing autonomous driving prototypes also cannot meet these requirements.
5) Cost control faces challenges. The current cost of autonomous driving system prototypes does not meet the cost requirements for industrialization.
The aforementioned problems are essentially due to insufficient depth and breadth of data accumulation in intelligent driving vehicles, a lack of strong computing power, poor task adaptability, and difficulties in optimizing and adapting AI algorithms. To solve these problems and achieve deep integration of AI in in-vehicle terminals, a vehicle-cloud collaborative intelligent driving system is considered. Leveraging the flexible and abundant computing resources of the cloud platform, complex AI algorithms can be processed, and the analysis results can be sent to the vehicle for real-time decision-making and planning. The cloud domain, acting as a network-enabled open brain and core, becomes the link connecting the internal network and the vehicle's business needs, thereby truly realizing network intelligence. Based on the development of cloud computing and big data technologies, the autonomous driving system is divided into two layers: vehicle and cloud (platform), proposing a vehicle-cloud collaborative autonomous driving system architecture. The cloud provides data storage, data sharing, and computing resources, supporting complex AI algorithms such as deep learning, autonomous learning, autonomous maintenance, and personalized adaptation. Through partial software/hardware sharing technology, vehicle-side costs can be reduced, computational load can be decreased, and the development of embedded AI hardware products for vehicles can be facilitated to meet the requirements of automotive-grade components.
2. AI-based vehicle-cloud collaborative autonomous driving system architecture and key technologies
Focusing on the three key elements of AI technology in autonomous driving—data, computing, and algorithms—and addressing the needs of multi-vehicle, multi-scenario, and personalized intelligent driving, this paper proposes an AI-based vehicle-cloud collaborative autonomous driving system architecture, as shown in Figure 3, to tackle the problems faced by single-vehicle intelligent driving systems.
Figure 3. Schematic diagram of AI-based vehicle-cloud collaborative autonomous driving system architecture.
The architecture consists of two parts: AI-based autonomous driving intelligent vehicle-side equipment and a cloud-based autonomous driving system based on big data analysis. Together, they form a vehicle-cloud collaborative autonomous driving system that integrates accurate perception of complex environments, intelligent traffic decision-making, and optimized execution of driving control.
2.1 AI-based autonomous driving intelligent terminal
An autonomous driving intelligent terminal is a cyber-physical system (CPS) integrating multiple functions such as environmental perception, planning and decision-making, and execution control. To adapt to the application needs of autonomous driving in different scenarios and vehicle models, it is necessary to conduct in-depth research on the hardware and software co-design technology of embedded intelligent controllers for autonomous vehicles. This involves establishing a hardware and software architecture for intelligent terminals that can support sensor data acquisition, environmental perception data fusion, planning and decision-making, and execution control AI algorithms, meeting the requirements of autonomous driving. The design of a real-time reliable autonomous driving AI terminal with system fault tolerance and "limp-in" capabilities is crucial. Furthermore, it is essential to propose real-time reliable, task-adaptive dedicated system software for intelligent terminals, enabling system integration verification and real-vehicle application of AI algorithms. Key technologies requiring breakthroughs include a real-time reliable autonomous driving AI terminal hardware architecture, a reliable and adaptive autonomous driving AI terminal software architecture, and the integrated application of AI technologies in autonomous driving intelligent terminals.
1) Hardware architecture of autonomous driving AI terminal
An autonomous vehicle AI terminal is a comprehensive intelligent system integrating multiple functions such as environmental perception, planning and decision-making, and control execution. Based on the different task divisions, working modes, and communication interconnection methods exhibited by autonomous driving systems in typical application scenarios for business modules such as environmental perception, planning and decision-making, and execution control, this study investigates the system reliability design and modular design methods of autonomous driving AI terminals. The focus is on the heterogeneous multi-core hardware system architecture based on GPUs and MCUs, and the high-speed interconnection communication architecture based on Ethernet.
2) Autonomous Driving AI Terminal Software Architecture
Autonomous vehicle systems integrate multiple software functional modules (environmental perception, planning and decision-making, execution and control, navigation, positioning, traffic signal monitoring, etc.) and multiple hardware execution units (computing units, control units, sensors, etc.). Research:
AI-based functional application software system architecture and hierarchical, modular design methods for perception, planning, and execution;
Optimal architecture for task-adaptive system and application software;
Ensure reasonable allocation and scheduling of hardware and software resources, including GPU, CPU, memory, bus and communication interfaces, and provide system self-healing capabilities, module resource isolation capabilities, computing and memory resource allocation capabilities, priority execution capabilities, and effective communication capabilities between modules.
3) Technical integration and application of autonomous driving AI terminals
As a typical physical-information fusion system, autonomous driving systems must utilize AI methods to achieve comprehensive integration of data and knowledge information.
To address the limited hardware and software resources of autonomous driving intelligent terminals, an AI operating system is being developed specifically for these terminals, enabling real-time execution of tasks such as perception fusion and decision-making control. In addition to possessing all the functions of a general-purpose operating system, the AI operating system should also include speech recognition, machine vision, actuator systems, and cognitive behavioral systems, and can be divided into an infrastructure layer, a technology development layer, and an integration and application layer. AI-based autonomous driving intelligent terminals have already received widespread attention in the industry, and numerous AI technologies are being applied to the field of autonomous vehicles at an astonishing pace.
However, some problems still need to be addressed:
For example, AI algorithms require a large number of labeled sample libraries for self-learning, and their underlying mechanisms are unclear and their boundary conditions are uncertain.
The application scope of AI technology is limited by factors such as the processing capabilities of automotive chips and sensors.
Therefore, it is important to upgrade sensors in sync with the automotive industry to improve data acquisition quality and provide hardware-level solutions for data fusion; and to maximize the application of AI technology in specific scenarios, such as closed/semi-closed areas, low/high speed conditions, rail transit, and special vehicles.
2.2 Cloud-based autonomous driving system based on big data analytics
With the powerful computing capabilities of cloud computing platforms, and considering the needs of multiple vehicle models, multiple scenarios, and personalized driving, this paper analyzes the requirements of AI in autonomous driving systems regarding data quality and access efficiency.
Research on AI-oriented cloud computing platform data space construction technology to achieve normalization of multi-type and multi-domain data on both vehicle and cloud ends;
This research focuses on the information and data interaction and collaboration technology between the vehicle and cloud ends in autonomous vehicle systems. It aims to construct an information and data interaction and collaboration framework, solve the problem of seamless connection between information and data on both ends of the vehicle and cloud, and complete the information and data subscription on the vehicle end and the information and data distribution on the cloud end.
Based on this, we will study the adaptation of AI algorithms to various scenarios under different vehicle models and driving styles, so as to enable autonomous vehicles to exhibit deeper intelligence at the three levels of perception, decision-making and execution, thereby improving the overall intelligence of autonomous vehicles.
1) Cloud Data Space Construction Technology
To reduce the complexity of AI data processing and information services in autonomous driving systems across multiple scenarios, vehicle models, and personalized driving environments, this study analyzes the distribution, heterogeneity, time-varying nature, and massive volume characteristics of the information data.
This research focuses on metadata description methods based on information data sources, metadata conflict reduction techniques, and metadata publication and discovery techniques to achieve the construction and management of metadata datasets.
Research the organizational structure and modeling techniques of information data space, and construct the object association set of information data space;
This research focuses on indexing and retrieval technologies based on metadata entity objects to enable the publication and discovery of heterogeneous information data sources based on metadata.
2) Vehicle-cloud collaborative technology
In different driving conditions and application scenarios, whether it is online AI learning and training for autonomous driving or offline interactive information preparation, a large amount of information and data interaction and collaboration are required between the vehicle and the cloud to implement accurate driving environment perception, intelligent traffic decision-making and optimized driving action control.
The vehicle-cloud collaboration technology for AI-based autonomous driving systems primarily addresses the issue of unified and effective data transmission between the vehicle and the cloud platform. The sampling data from the vehicle's sensor nodes includes numerical data (such as GPS/INS data and millimeter-wave radar data) and multimedia data (such as camera images). This sensor data is transmitted to a cloud database at a certain frequency for online processing, offline processing, traceability processing, and complex data analysis.
The autonomous driving system proposed in this paper includes two intelligent subsystems: the vehicle-side and the cloud-side. The cloud-side system can not only store massive amounts of real-time sensor data, but also store historical data. It also uses cloud computing to store, transmit, analyze, and process this massive amount of data. Based on the intelligent driving control model with integrated AI application algorithms, it provides a reliable and efficient collaborative control scheme for vehicle decision-making.
3) Cloud platform AI algorithm application technology
Cloud platform AI algorithm applications are a core component of autonomous driving cloud systems. They combine machine learning, data mining, and other related technologies to analyze perceived and fused information, providing decision-making support for vehicle control planning. However, in-vehicle embedded hardware platforms, due to their limited computing and storage capabilities, cannot meet the training requirements of AI models. Autonomous driving cloud platform AI algorithm application technology utilizes virtualization and networking technologies to integrate large-scale, scalable distributed computing resources, including computing, storage, data, and applications, to complete the learning and training of AI model algorithms. This enables AI models to be trained in the cloud and deployed to embedded platforms through vehicle-cloud collaboration technology, allowing for the deep application of AI algorithms in in-vehicle autonomous driving systems.
It is foreseeable that the main challenges facing autonomous driving cloud systems in the future will be concentrated on ultra-large-scale data storage, data encryption and security assurance, and improving I/O speed; in terms of technology implementation, they will also face limitations in supplier collaboration and operational pricing policies.
Therefore, considering fully utilizing existing innovation resources and platforms, integrating data from various enterprise-level platforms and government regulatory platforms, transforming the closed nature of communication and publishing systems, and adopting a cloud computing model to build a transportation service system, some small and medium-sized cities only need to rent the corresponding services, which is conducive to the popularization of autonomous driving cloud systems. At the same time, with government guidance, core drafts such as cloud platform technical specifications and data element formats should be compiled and promoted, thereby guiding the demonstration, promotion, and sustainable development of autonomous driving cloud system applications.
3. Conclusion
This paper analyzes the current development status of autonomous driving technology, as well as the application trends and challenges of AI in autonomous driving. Based on this analysis, an AI-based vehicle-cloud collaborative autonomous driving system is proposed, and its system composition and key technologies are described.
The article proposes that the deep application of AI in the field of autonomous driving requires key solutions to problems such as deep integration of AI in autonomous driving vehicle terminals, data normalization of autonomous driving system vehicle cloud computing platform, vehicle-cloud information data interaction and collaboration, and vehicle-cloud AI algorithm adaptation to multiple vehicle models and scenarios as well as personalized driving.
For applications involving multiple vehicle models and scenarios, it is pointed out that research is needed on the hardware and software co-design technology of embedded intelligent controllers for autonomous vehicles.
To address the limited local storage and computing capabilities of in-vehicle autonomous driving systems, this paper proposes using cloud computing as an extension of vehicle-side capabilities to solve the problems of large data storage space and HPC capabilities necessary for AI algorithm model learning and training.
To address the interaction issues between the vehicle and the cloud, a vehicle-cloud collaborative approach is proposed to deploy AI models trained in the cloud to the vehicle for execution, enabling autonomous driving tasks such as perception fusion and planning decision-making.
Ultimately, a complete vehicle-cloud collaborative autonomous driving system is formed based on the vehicle-side and cloud-side hardware and software architecture.
