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

With the improvement of vehicle automation level, in order to obtain richer and more accurate environmental information, vehicles are usually equipped with multiple sensors. Environmental perception technology based on multi-sensor information fusion is currently a research hotspot. Information fusion algorithm is the core of realizing multi-sensor environmental perception technology. In view of the problem that a single sensor in the current vehicle collision avoidance system cannot effectively identify the vehicle target in front, reference [64] uses joint probability data association to fuse radar and camera data. Reference [65] models the complex traffic scenes obtained by lidar and camera, and uses hybrid system theory to obtain the optimal acceleration of the vehicle, which is used to effectively enhance the adaptive cruise control performance in a multi-vehicle target environment. Reference [66] integrates monocular camera, lidar, millimeter wave radar, ultrasonic radar and other sensors, and designs an optimal Gabor filter through machine learning algorithm for obstacle detection, identification and classification.

Environmental perception system is one of the key links in the development of automobile driving automation. Its perception ability directly determines the level of intelligence of smart cars. The current sensor layout scheme and perception technology can ensure effective perception under specific working conditions. However, for complex road environments such as roundabouts, intersections, unstructured roads, and strong interference environments such as lighting, noise, buildings and weather, the accuracy of perception and information fusion is difficult to guarantee. The main problems are analyzed and summarized as follows: 1) The robustness in complex traffic environments and bad weather needs to be strengthened, such as pitch/roll on unpaved roads, extreme weather such as rain and snow, and flickering light when entering and exiting tunnels. 1) The interference of various factors will have a great impact on the accuracy of the perception system; 2) When the vehicle is driving at high speed transiently, it is difficult to accurately obtain information such as the relative speed and relative acceleration between the vehicle and the target, which makes it difficult to fundamentally verify the accuracy of the perception system; 3) The accuracy of the target-level information constructed from the original-level measurement data in complex traffic environments still needs to be improved; 4) It is not clear how the target-level information obtained by the perception system is distributed to each control subsystem on demand. In addition, the control system related to automobile safety requires the perception system to provide rapidly updated perception fusion information, and the rapid implementation of advanced information fusion algorithms is still a problem that needs to be solved urgently.

2.2 Insufficient autonomous planning and decision-making capabilities to adapt to open and uncertain environments

In an autonomous driving system, decision-making and planning are the central command system, equivalent to the brain of the vehicle. After receiving various sensory information, the decision-making and planning system combines the analysis of the current environment and vehicle status to issue instructions to the underlying execution control system. In autonomous driving vehicles, decision-making and planning are generally divided into three levels: path planning, behavior decision-making, and motion planning. For the current development status of autonomous driving decision-making and planning, please refer to reference [67]. The following is only a brief overview.

Path planning can generally be divided into global path planning and local path planning. Global path planning is to plan selectable paths from a macro level based on the starting point, end point, and road topology. Local path planning is to fully consider the feasibility of the path on the basis of global path planning[68] and further refine it. The main methods for path planning include grid method (graph search method), sampling method, curve interpolation method, numerical optimization method, etc.[69]. The behavior decision layer makes specific optimal behavior decisions for the vehicle after receiving the planned path, such as following, changing lanes, and overtaking, in combination with the environmental information of the perception system[70]. Its main methods include rule-based[71,73], learning method[72], random theory[74], swarm intelligence algorithm[75], etc. After the decision behavior is determined, the process of converting the behavior into a specific driving trajectory is motion planning. Motion planning has a long research history in the field of robotics[71] and has gradually been transplanted to autonomous driving.

The ability of decision-making and planning systems to handle scene complexity is one of the core indicators for measuring and evaluating autonomous driving capabilities. Currently, decision-making and planning systems are mainly used in simple working conditions or closed road environments. In actual complex working conditions with various uncertainties, their ability to make reliable autonomous decisions is still insufficient. The main problems are summarized as follows: 1) Autonomous driving decisions in open scenarios need to consider many uncertainties. These uncertainties mainly come from perception errors [76], model errors, obstacles [78-79], and the uncertainty and randomness of the behavior of other traffic users [77]. How to predict and estimate the uncertainty in the environment is still a difficult problem; 2) At present, autonomous driving Path planning for autonomous driving needs to consider a variety of motion constraints. How to plan a smooth path that meets the vehicle's kinematic constraints while meeting the algorithm's real-time computing speed requirements is an urgent problem to be solved; 3) Compared with the driver's decision, the current autonomous driving decision-making system cannot "think" about problems like human drivers. Taking lane changing decisions as an example, autonomous driving may make "stupid" decisions to change lanes repeatedly and frequently under safe conditions. In short, analogous to human driving decisions, the autonomous driving decision-making system is not yet perfect in the development of various aspects of human-like decision-making such as advance avoidance[80], game[81], memory[82] and uncertainty estimation[83], and further exploration is needed.

2.3 Coordinated vehicle motion control has not yet been achieved

With the increasing degree of automation of automobiles, the integration complexity of their motion control systems is also increasing (including control subsystems in the vehicle's side-to-longitudinal-vertical directions, such as TCS, ESC, ABS, etc.). In current practical engineering, vehicle motion control usually allows one subsystem to intervene while another subsystem exits. The action boundaries of each subsystem are calibrated through rule-based methods, and the coordination between systems has not been fully realized. Under extreme conditions (referring to the "unconventional conditions" where the vehicle dynamics system enters the nonlinear zone, as shown in Figure 3), there are obvious interferences and conflicts between the motion control subsystems and related actuators. How to coordinate and expand the vehicle's stability boundary is a difficult problem. It is mainly reflected in: 1) There is complex nonlinear coupling in the side-to-longitudinal-vertical dynamics, saturation of tire forces, and tire side 1) The characteristics of partial stiffness are time-varying, and there is coupling between the safety control objectives of each subsystem; 2) In the context of the development of autonomous driving, active safety control needs to coordinate the upper-level planning and decision-making, and the calibration quantity of each subsystem has increased dramatically. In order to meet the functions of accurate and fast execution of upper-level autonomous driving, each subsystem must have good control performance. However, the current motion control systems are developed separately, and the distributed control scheme has brought obstacles to the coordinated execution of planning, decision-making, and control. Therefore, in order to fundamentally solve the interference and conflict problems between subsystems and actuators and expand the stability boundary under extreme working conditions, it is necessary to start from the top-level control system architecture design and realize the integrated collaborative control of active safety of vehicle motion by directly optimizing the lateral-longitudinal-vertical dynamics.

Fig.3 Schematics of extreme driving condition (lateral-longitudinal )

At present, vehicle cooperative control is still in the scientific research stage. Reference [84] proposed a cooperative control strategy for the subsystems of active front and rear wheel steering and braking/driving force distribution. The wheel load balancing based on braking/driving force distribution can improve the steering performance of the vehicle. At the same time, a yaw moment compensation method is used to reduce the yaw rate tracking error. The entire cooperative control strategy can achieve load balancing and rapid tracking of the desired yaw rate. References [85−88] respectively use constrained H∞ control methods and robust nonlinear control methods to deal with dangerous conditions of tire blowouts of high-speed vehicles. By coordinating the steering subsystem and the braking subsystem, the trajectory tracking control and directional stability control of the tire blowout vehicle are achieved. Reference [89] discusses the cooperative control of four-wheel steering angle and active suspension adjustment. The method controls the active suspension force under the condition of small tire slip angle, thereby improving comfort; when the tire lateral force reaches saturation, the tire slip angle is adjusted by controlling the comprehensive means of the vehicle's four-wheel steering angle and vertical force to achieve the vehicle stability control target. Reference [90] aims to improve the handling stability under extreme conditions. Through the integrated control of vehicle dynamics, fuzzy logic rules are used to distribute the four-wheel torque and rear wheel steering angle to improve the vehicle's handling performance and prevent the vehicle from deviating from its trajectory on slippery roads. The existing collaborative control mainly studies a single safety target under typical conditions. The problem of coordinated optimization control of nonlinear coupling systems under extreme conditions is still an open problem. How to use network information to improve the active safety performance of vehicles is also a difficulty today.