A 10,000-word article on automobile control in the intelligent era

The operational control of the mechanics demonstration experiment includes three characteristics: first, the demonstration operation of the teacher must be correct and standardized, so as to ensure the success of the demonstration phenomenon and allow students to develop good operating habits when watching the teacher's operation; second, according to the purpose of the experiment, classroom needs and students' reactions, the teacher should effectively control the speed, number of times, direction, position, etc. of the demonstration experiment process, and give students appropriate language inspiration and guidance during the operation process; finally, without affecting the phenomenon of this mechanics demonstration experiment, certain special controls are performed on the operation process to make the demonstration phenomenon more intuitive and obvious, which is convenient for students to analyze and summarize the experimental conclusions or experimental laws, thereby achieving the purpose of this demonstration experiment.

2.4 Human-machine collaborative driving is just getting started

In addition to the coordination problem between vehicle motion control systems, there is also interaction and coordination between the driver and the vehicle, and the difficulty of achieving human-machine collaboration is greater. That is, how to coordinate the weights of the human-machine hybrid driving decision-making automatic control system and the human driver to form a dynamic human-machine interaction relationship and achieve optimal allocation and switching of driving rights between the driver and the control system.

At present, human-machine collaborative driving technology is still in the scientific research stage. Reference [91] uses a model prediction method to use road information to plan an optimized path to guide and help drivers drive better. Reference [92] improves the Copilot concept and gives suggestions on the degree of automation. The driver makes a comprehensive decision based on multiple factors, so the driver has direct control. Since the control of this type of system is determined by the driver, when the driver is not paying attention or ignores the guidance given by the automatic controller, driving errors will occur, and it is difficult to effectively avoid traffic accidents. In order to avoid traffic accidents, some researchers use a driver and automatic controller switching control method. This method analyzes road information and driver information to determine whether it is suitable for human driver control, and switches the driving rights between the human driver and the automatic controller. Reference [93] switches between driver control, assisted driving and automatic driving modes through situation analysis and risk assessment. In order to solve the stability problem of frequent switching of human-machine co-driving systems, reference [94] uses a composite Lyapunov The function analyzes the stability of the three modes of switching: single-driver driving, assisted driving, and automatic driving. The optimization of human-machine collaboration and the decision on control rights are based on the interactive learning between the driver and the machine. If the vehicle dynamics enter the nonlinear region, ordinary drivers are prone to improper operation due to lack of experience and psychological panic, causing the car to become unstable. Therefore, how to consider the complexity of human-machine coupling and the driving environment, the driver's driving habits and behavioral uncertainty, is a challenging problem faced by human-machine co-driving. For more discussion on the current status of human-machine co-driving research, please refer to reference [95], and this article will not go into too much elaboration.

3 Development Trends and Key Technologies of Automobile Control in the Intelligent Era

The rapid development of emerging technologies in the intelligent era has provided new opportunities for the development of automobile driving automation. The continuous development of ubiquitous sensing (such as high-precision maps and network technology) can provide automobile control systems with richer, faster and more accurate environmental information; the rise of technologies such as big data analysis and artificial intelligence under ubiquitous computing has provided more intelligent and efficient means for automobile control and optimization decision-making. In short, the application of these new technologies in automobiles has brought all the performance of automobiles to a new level, making the vehicle control system with automobiles, environment and drivers as carriers gradually become a typical integrated intelligent system (Cyber-Physical-Human Systems) integrating information resources, physical resources and human activities. The theme of automobile control in the intelligent era is to integrate emerging technologies in the intelligent era into the perception, cognitive modeling, intelligent decision-making and execution of automobile control systems, and realize new control systems and new functions to improve the safety, energy saving, environmental protection, comfort and economy of automobile control systems. The following mainly analyzes the future development trend of automobile control in the intelligent era from six aspects: on-board computing and communication, big data information fusion, application of advanced control theory in automobile control, new functions of intelligent automobile control systems, human-in-the-loop automobile control, and testing and evaluation of automatic driving.

3.1 Automotive Control and Onboard Computing and Communication Technologies in the Intelligent Era

With the increasing degree of automation, cars are gradually becoming data interaction and computing centers on wheels. Traditional on-board computing ECUs (Electronic control units) will find it difficult to meet the future needs of highly automated vehicle control. On-board edge computing platforms that take into account multiple requirements such as speed, safety, scalability, and low cost will be a rigid requirement for automotive control in the intelligent era. Field programmable gate arrays (FPGAs) with multi-threading and parallel processing capabilities are a feasible solution for the rapid implementation of advanced automotive control and optimization algorithms[96]. At the same time, computing platforms based on remote computing clusters or on-board workstations[97] are solutions for the conceptual development and testing phase of autonomous driving control technology, but their disadvantages such as high energy consumption and high cost are still challenges for mass production. NVIDIA has developed the Drive PX Xavier high-performance computing chip for autonomous driving[98], and Mobileye has developed the EyeQ series of chips. Its EyeQ5 is expected to be released in 2020 and claims to have a maximum computing power of about 10 watts[100]. Google released TPU in 2016, a chip designed for Google Designed for the deep learning platform Tensorflow framework, it can achieve up to 30 times the performance improvement and up to 80 times the efficiency improvement compared to the GPU of the same period[101]. In July 2018, Baidu released the all-round AI chip "Kunlun" and claimed that it was the computing chip with the highest computing power in the industry. Qualcomm successfully transplanted the "Snapdragon" chip from the mobile phone platform to the vehicle platform, integrating 5G communication, neural network processing engine, GPS, WiFi and other multi-functional chips on a circuit board, which can meet the needs of future intelligent network and highly automated vehicles. At present, whether it is vehicle manufacturers, Internet companies or universities, they are investing huge manpower and material resources to seize the strategic high ground of intelligent vehicle computing platforms. Research on fast, reliable, secure and low-cost edge computing solutions will always be a hot spot for the future development of intelligent vehicles. In addition, centralized or distributed high-performance computing resources such as cloud computing and fog computing[99] will also provide support for the development of intelligent vehicles in the future.

In the process of evolution from driver assistance to advanced autonomous driving, images, laser point cloud data, and a large amount of network/sensor information have gradually penetrated into the control system, which has also put forward higher requirements for vehicle communication technology. Currently, there are two main wireless communication technologies, namely dedicated short range communication (DSRC) and 4G LTE[102]. However, neither of these two technologies can provide gigabit/second data rates, high-speed mobility support, large-scale machine communication, and ultra-low latency, and it is difficult to meet the requirements of fully autonomous driving. Automobile control in the intelligent era is also one of the important application scenarios of 5G technology. Compared with 4G communication technology, the data transmission speed of 5G can be increased by 100 times, and the end-to-end delay can be as low as milliseconds[103]. Therefore, 5G wireless communication technology will bring powerful support to future fully autonomous driving systems and bring huge improvements to various core performance indicators such as automobile energy saving, emissions, and safety.

3.2 Multi-source heterogeneous big data information fusion technology

The lack of perception ability in complex environments is the bottleneck that restricts vehicles from achieving fully autonomous driving. In the intelligent era, cameras, radars, GPS, high-precision maps, and connected environments (V2X) will provide a large amount of multi-source heterogeneous information for the vehicle's perception system [104]. Whether the information comes from on-board sensors, roadsides, or surrounding vehicles, information fusion technology is essential. Emerging multi-source information fusion technology based on artificial intelligence can help understand the complexity of the environment and the inherent correlation between information, expand the perception space and time range, and improve the credibility of perception, thereby greatly improving the perception ability and robustness of the vehicle. It is the development trend of autonomous vehicle perception technology. In the future, fusion algorithms will continue to be a hot topic in multi-sensor data application research. Traditional methods represented by the Kalman filter algorithm rely on recursive data fusion [105-106]. Fusion algorithms based on artificial intelligence mainly include neural networks and deep learning. [107] aims to design a network topology structure, take multi-source information as input, learn and understand the output information, determine the network weight distribution, and complete information fusion and knowledge acquisition. For low-cost navigation systems, reference [108] proposes a neural network method to fuse inertial navigation and GPS signal measurement data to obtain accurate location information. Deep learning methods generally use supervised deep learning, which requires the use of labeled data for training, but the workload of manual labeling is huge. Unsupervised learning algorithms do not require labels, but require continuous trial and error to implement, which is more risky. Reference [109], in the field of target classification based on multi-image feature information perception, in order to improve the training efficiency of convolutional neural networks in deep learning, based on the design of a local recommendation network, combined with the convolutional features of the shared image of the target detection results, proposed a learning-based information fusion method, which to a certain extent solved the space overhead when training massive samples.