A review on key challenges in intelligent vehicles: Safety and driver-oriented features

The huge advantages of intelligent vehicles (IVs) in improving road safety and operating efﬁciency have become a research focus in the industry. IVs have made signiﬁcant progress in recent years, but it is still face great challenges in order to be accepted by users on a large scale. In this regard, the authors propose that the research of IVs can be developed along the lines of safety, comfort and economy, gradually overcoming existing dilemmas. First of all, security is the most basic requirement of IVs. The authors sort out the key technologies and challenges of the basic architecture of IVs, and propose existing attack and defences strategies for IVs information security technology. Secondly, comfort is more about people’s subjective feelings. From two aspects of physiological comfort and psychological comfort, the paper studies the anthropomorphic decision-making to overcome the mechanized speed control, human computer interaction design, personalized driving style and ethical decision-making methods. Then, aiming at the micro-and macro-levels of economy, it outlines technologies such as economic driving behavioural, collaborative control of people, vehicles and roads and IVs sharing. Finally, the authors summarized the challenges and future development directions in the three stages of IVs development.


INTRODUCTION
With the rapid economic growth, the number of motor vehicles and the number of traffic accident deaths in China has become the first in the world. According to the data analysis of China Statistics Bureau, in 2019, China's automobile traffic accident rate dropped by 4.5% year-on-year, but there were still 159,335 automobile traffic accidents, resulting in 43,413 deaths and 157,157 injuries, causing huge economic losses [1,2]. And in the National Highway Traffic Safety Administration (NHTSA) report of the United States, 94% of the accidents are caused by driver's human error [3], and related to the driver's age and dangerous driving behavioural [4,5]. Studies have shown that intelligent vehicles (IVs) have great potential in improving road safety, passenger comfort, road traffic efficiency, energy conservation and emission reduction [6]. Therefore, the field of IVs has become the research focus of scientific research teams all over the world, and has achieved remarkable results. Secondly, according to the data of the World Health Organization, the number and proportion of the global elderly population continue to grow, and the scale is expected to reach about 2 billion by 2050 [7]. The aging trend of the world population is irreversible, and a series of traffic problems will occur at the same time. In the event of an accident, older drivers are unable to respond quickly in a short period of time, especially when cognitive function begins to decline. Simplifying operation tasks helps to improve safety [8]. In addition to the elderly, there are also vulnerable groups with difficulties in driving vehicles, such as people with disabilities, children and women. In summary, the development of IVs plays an important role, not only to meet the social interests, but also to simplify the way of travel.
Although the development of IVs has many benefits, there is still a long way to go for large-scale application. First of all,  [10][11][12][13] The literature in one box represents the same main scope.
as the most basic requirement of IVs, safety has always been a research hotspot, especially in recent IVs accidents, we have to re-examine this topic. Secondly, comfort is to consider human factors on the basis of safety. Studies show that most people feel uncomfortable and uncomfortable with IVs. They are afraid of the potential danger of losing control [9]. Thirdly, the economy (Energy Saving, Environmental Protection) is also an important factor affecting the landing of automatic driving under the condition of meeting the user's safe and comfortable experience. The safety, comfort and economy of IVs interact with each other and are indispensable. Therefore, it is necessary to study the development mechanism of these three factors. In addition, the existing literature review is relatively small, and most of the literature coverage is relatively narrow, mainly in environmental perception, decision control and other aspects, the specific comparison is shown in Table 1. This paper aims to make a comprehensive and systematic overview of the research status in the field of IVs, so that we can have a new understanding. This paper is divided into five parts, the rest of the structure is as follows: In Section 2, we summarize the security framework of IVs, and discuss the communication security technology of IVs. In Section 3, we study the comfort and related concepts of IVs, including riding comfort and psychological comfort, focusing on anthropomorphic driving, human-computer interaction, moral and ethical decision making. In Section 4, we study the economic related technologies, including economic driving behavioural, IVs sharing, intelligent network coupling queue and so on. In Section 5, we make a systematic summary of the status quo of IVs development research, and look forward to the possible future development direction in view of the existing challenges.

SAFETY OF IVS
Safety [29] is the foundation of IVs landing and the most basic requirement of users. The safety problems of IVs mainly include functional safety, safety of the intended functionality (SOTIF) The basic architecture and functionality of IVs safety and information security. As a supplement to functional safety, SOTIF emphasizes to avoid unreasonable risks due to the limitation of expected functional performance. The basic architecture of IVs is an important part to solve the problem of safety, mainly including: environmental perception, intelligent decision making and control execution [19]. The specific architecture and functional safety issues are shown in Figure 1. In this section, we will discuss the research status of the above contents, as well as the existing challenges and future trends.

Environmental perception system security
The environmental perception technology is equivalent to the driver's eyes and ears, and needs to perceive the external The main task of environmental awareness information in real time. Its main task is shown in Figure 2. In the early stage, the environmental sensing system completed the detection task through various sensor systems, such as Vision system, Radar system and Lidar system. Due to the single sensor's own parameters and external environment interference, the sensing is limited (Field of view, Range, Direction, Number of scanning rays, Weather etc.), which cannot provide reliable 360 • environment sensing [30]. Single sensor often has insufficient perception, resulting in false detection and missed detection. The sensing range of environmental sensing system is expanded by fusing with other sensors [31,32]. In the sensor fusion system, the selection and combination of sensors directly affect the reliability and robustness of environmental perception of IVs, as well as vehicle production costs. Sensor fusion mainly includes radar and lidar fusion, radar and vision fusion, lidar and vision fusion and a variety of sensor fusion. Through the Bayesian theory, Kalman filter and Dempster-Shafer (DS) evidence theory, the fusion algorithm can accurately and reliably describe the external environment. In order to improve the perception accuracy of unsafe scenes in SOTIF, the development and verification of fusion algorithms in complex scenes are further strengthened based on the existing sensor fusion algorithms. Although sensor information fusion and redundancy design can effectively solve the problem of sensor failure, recognition algorithm error recognition and unrecognized, it still needs to solve the problem of motion compensation, time synchronization and real-time requirements [32]. Secondly, the existing sensor technology only provides the vehicle with the ability of 'seeing', while the remote communication technology and V2X technology can provide interactive perception to make up for the constraints of environment and distance. Jung et al. [33] proposed an over the horizon sensing system based on V2X communication technology. Through data fusion with the sensors of the vehicle, it can realize all-weather and uninterrupted accurate sensing. The application of 5G network technology in V2X communication can speed up data transmission and data security, and reduce the perception delay and instability characteristics [34].

Security of intelligent decision system
Intelligent decision system is the brain of IVs, which is mainly composed of global planning, behavioural selection and local planning [19]. Its main research framework is shown in Figure 3. Common global planning algorithms include Dijkstra and A* algorithms, as well as various improvements based on these two algorithms. To deal with the complex urban road network and ensure the safe and effective arrival of the destination, the classic path planning algorithm is obviously not enough. Through V2X communication technology, multimodal route planning considering traffic congestion, public safety, traffic management and weather factors is the future development direction [35].
In order to generate obstacle free trajectory, behavioural decision making needs real-time risk assessment of the surrounding environment and relevant road traffic rules to select the driving mode. The application scenarios of time index and dynamic index as early risk assessment indicators are relatively simple. Wang et al. [36] proposed the concept of 'driving risk field', considered the influence of various traffic elements in the closedloop system composed of people-vehicle-road on driving risk, and predicted the driving safety trend through dynamic changes. Gao et al. [37] use stochastic environment model and Gaussian distribution model, not only can accurately assess and predict the risk within the prediction range, but also can assess the risk of the scene outside the prediction range. Katrakazas et al. [38] constructed a joint risk assessment framework based on interactive perception motion model and dynamic Bayesian network (DBN). In risk assessment, other traffic participants' movements are predicted to calibrate the assessment results. The vehicle movement on the road is generally regular, and we can predict the vehicle's next trajectory based on the deep learning (DL) model (Gaussian Mixture Model, Hidden Markov Model) based on the steering, acceleration and deceleratio and distance from the lane [39]. However, pedestrians can quickly change their speed and direction because of their agility. The accuracy and real time of intention estimation are both very challenging. Ahmed et al. [40] proposes a DL to estimate the future intention of pedestrians, taking into account the dynamic motion model (DMM) of motion trajectory, the skeleton characteristics of pedestrians. In the previous method, pedestrians are regarded as independent individuals. Vemula et al. [41] proposed a social attention trajectory prediction model, in which pedestrians adjust their own trajectory according to the movements of other people around them. Finally, according to the behavioural decision results, the local planning chooses the optimal collision avoidance trajectory in limited time and under various constraints. In conclusion, the existing decision-making algorithms are mainly based on empirical rules, data-driven, utility function, interaction and uncertainty. However, these algorithms need a large number of calibration data, and the calculation is complex, resulting in low real-time performance, easy to fall into local optimal FIGURE 3 Intelligent decision system research framework solution and other functional limitations, leading to decisionmaking errors in unknown scenes. Through the establishment of a typical scene library which integrates various types, high complexity and uncertainty, and testing the robustness, adaptability and generalization ability of vehicle decision function to different scenes. On this basis, the decision algorithm is continuously optimized to improve the processing ability of IVs in SOTIF scene.

Control execution system security
The control execution system is the key to realize the autonomous driving of IVs, and the level of control is directly related to the safety of vehicles. The control module is equivalent to the hands and feet of the human driver, which is used to implement decisions to realize the lateral and longitudinal control of the vehicle. Its control framework is shown in Figure 4.
In the development of control algorithm, the longitudinal and transverse decoupling is often used under normal conditions. Zheng et al. [42] decoupled the longitudinal and lateral motion planning to realize the trajectory replanning in the normal lane changing process to avoid collision. In the face of complex traffic environment, single lateral/longitudinal control and simple coupling relationship will lead to weak robustness of the system. Aiming at the non-linear coupling relationship of vehicle lateral/longitudinal motion, the lateral dynamic model and longitudinal dynamic model of vehicle are integrated in the same FIGURE 4 Control execution system framework control framework, which has strong robustness [43]. At the same time, the advantages of model predictive control (MPC) are outstanding. It is not only simple in structure, but also can deal with complex process models with input constraints and non-linearity, and can specify the system input, state and output of the system [44]. In addition to the deviation from the ideal target caused by the limitation of its own model function, it also faces the SOTIF caused by the external environment, such as the limitation that the vehicle dynamic model is not enough to [45] used two-degree-of-freedom dynamic model to design linear quadratic regulator (LQR) algorithm controller, which can well solve the steady-state tracking error of curve driving. Ji et al. [46] proposed an IVs path tracking framework based on multiconstraint MPC considering geometric constraints of road and dynamic constraints of vehicle. Although the researchers have improved the existing models, the research on control execution under extreme conditions is not enough, such as the response ability of the system under the boundary conditions such as the minimum road adhesion rate, the maximum allowable lateral force interference, the maximum longitudinal slope, the maximum execution deviation and so on. Secondly, the failure control of the actuator itself is still a big problem, and the functional redundancy of the actuator provides an opportunity to reduce the safety requirements of a single actuator.

Communication and information security
More and more electronic control units and external communication technology interfaces are used in IVs, and the accompanying hacker attacks and network communication security are threatening the safety of vehicles and users' information privacy. In recent years, the occurrence of vulnerabilities in BMW's digital service system, remote Tesla intrusion incidents, and Nissan LEAF automobile API leaks have proved the necessity of information security research. The current sources of security risks mainly include external mobile communications, vehicle-mounted networks, vehicle-mounted terminals and cloud platforms, as shown in Figure 5. In order to accurately respond to information security threats, it is necessary to have a correct understanding of existing attacks.
There are many types of attacks that affect the communication security of smart vehicles. Dibaei et al. [11] divided the types of attacks into denial of service attacks, distributed denial of service attacks, black hole attacks, reply attacks, Sybil attacks, Impersonation attacks, malware, falsified-information attacks and Timing attack. Considering that the degree of intelligence and networking of IVs systems is constantly improving, there are multiple attack portals in the vehicle life cycle, and a secure system needs to detect the type of attack of the system in real time. Early machine learning methods are widely used to identify various types of attacks, but the accuracy will decrease with the increase of classification tasks. Yin et al. [47] established an intrusion detection system model based on DL, and proposed a DL method using recurrent neural network for intrusion detection. Experimental results show that Recurrent Neural Network-Intrusion Detection System (RNN-IDS) is very suitable for create classification models with high accuracy, and its performance is better than traditional machine learning classification methods in binary and multi-level classification.
In addition to DL as the main method of intrusion detection, there are signature-based detection, anomaly-based detection, malware detection and software vulnerability detection. As the number and complexity of intrusions increase, a single or isolated IDS is ineffective in many cases. In this regard, Meng et al. [48] designed collaborative intrusion detection systems/networks (CIDSs/ CIDNs) to allow intrusion detection system nodes to collect and exchange information required by each other. For the IVs communication system itself, the encryption and authentication features are effective countermeasures to reduce hostile attacks. Symmetric encryption, asymmetric encryption and attribute-based encryption are conventional methods, but there are still data leakage, high computational load and long delay time. To solve this problem, Ying and Nayak [49] proposed an anonymous lightweight authentication scheme based on the smart card protocol, which uses low-cost encryption operations to verify the legitimacy of the user (vehicles) and verify data messages. In the next stage, we can improve the accuracy and response time of the detection system, establish a unified communication protocol standard, optimize the computing resources of machine learning and obtain a data set that can be effectively trained. Secondly, blockchain technology was first applied to the cryptocurrency Bitcoin, which is an immutable peer-to-peer distributed database containing encrypted security information [50]. Considering the environment of IVs communication network and how to establish a reliable trust network based on blockchain between IVs will be another communication security revolution [51].

COMFORT OF IVS
The biggest challenge of IVs is not only safety, but also human factors. The comfort of IVs mainly considers the subjective feeling of passengers, including physiological comfort and psychological comfort [52]. From the level of physiological comfort, it is necessary to ensure that the impact of speed change on the user's body is within a certain limit. From the perspective Research framework of IVs comfort of psychological comfort, driving mode is selected according to users' preferences, and information interaction with other traffic participants is mainly involved in trust and ethics issues. The research framework of IVs comfort is shown in Figure 6.

Physiological comfort
The causes of autopilot diseases are mainly due to unstable speed, excessive acceleration and deceleration, unstable posture and sensory conflict [53]. In particular, IVs make users lose control and cannot predict the future trajectory, which promotes the incidence and severity of carsickness [54]. Secondly, it can reduce the workload of users during the ride, and personalized design can also improve the physiological comfort of users. Therefore, some anthropomorphic decision-making control and human-computer interface interaction are used to alleviate the impact of physiological comfort.

Anthropomorphic decision making
The comfort performance can be effectively improved by learning the decision-making control of skilled human drivers in complex and potentially dangerous situations. Guo et al. [55] proposed a method for local trajectory planning by generating mixed potential graph of anthropomorphic behaviour. The method uses Bayesian network to generate the trajectory induced potential energy of the surrounding environment, and considers the driving skills and experience of drivers in real traffic environment, as well as traffic rules. He et al. [56] proposed a new cost function, which considered the safety and comfort of the track, and mainly referred to the lane change decision of human drivers in the natural driving data for lane incentive.
Li et al. [57] proposed a social intelligent empirical decisionmaking network that imitates human beings to deal with the coexistence of human drivers and IVs on existing roads, and the misunderstanding between the two causes traffic conflicts and affects comfort.
The traditional path tracking method is always eager to correct the error between the planned trajectory and the current state of the vehicle, which makes the vehicle need to continuously conduct steering, braking and acceleration operations. This process makes intelligent driving appear abnormal and too cautious. In this regard, Wei et al. [52] proposed a vehicle motion control frameworks for risk corridor. The Nonlinear Model Predictive Control (NMPC) model was established by using lateral offset tolerance and existing vehicle dynamic constraints under the condition of considering the comfort and safety of passengers. The results show that the method can track the planning path with a smooth trajectory. Zhu et al. [58] proposed a depth deterministic policy gradient (DDPG) algorithm, in which the reward function was learned from the natural driving following data, and finally the optimal strategy or vehicle following model was obtained from the anthropomorphic mapping of vehicle speed, relative speed between front and rear vehicles, distance between vehicles and acceleration of rear vehicles.
In addition, the speed control strategy can effectively improve the comfort. Du et al. [59] proposed a disturbance rate model to describe the individual vibration sensitivity. The theoretical speed was calculated to maximize passenger comfort, and the annoyance rate was used to modify the evaluation results. González et al. [60] use Bezier curve to smooth acceleration and bump curve, and improve riding comfort of automatic driving vehicle by limiting global acceleration in the whole driving process. In order to ensure the vehicle comfort when the vehicle speed changes significantly, Wu et al. [61] proposed an adaptive cruise control (ACC) system with an active braking algorithm, and an upper decision controller based on the MPC algorithm. The results show that the speed and distance of the vehicle are always within the specified comfort range.

Interaction decision
As human drivers get rid of the control of vehicles, humancomputer interaction becomes more important, mainly divided into internal interaction and external interaction [62]. With the change of user identity, more and more demands are needed. Human-computer interaction as a link of information transmission with IVs system. In order to meet the interaction needs of a variety of people and reduce user cognition, the multi-channel fusion of multiple sensory channels (Visual, Auditory, Smell, Touch, Taste, Body Feeling) of human beings is integrated to generate interaction with the system [63]. Research by Manawadu et al. [64] has proved that multi-modal human-computer interaction can achieve comfortable riding experience by promoting efficient interaction and reducing user workload. In order to further improve the physiological comfort, the intelligent cockpit has a powerful situational awareness system, which judges the user's heart rate, respiratory rate, age, gender, shape etc. through the biosensor technology, so as to provide users with different scene riding experience [65]. Differences in users' visual information, sense of balance and expectation may also cause physiological discomfort [66]. In this regard, Sawabe et al. [67] raised a reduction of reality method based on acceleration stimulation, which can reduce the motion sickness of IVs by showing the intention of the vehicle to the user before the actual acceleration occurs and guiding the passenger's centre of gravity to move. Wang et al. [68] proposed a vehicle collision pre-warning algorithm based on driving safety field model. The algorithm can effectively express collision risk in various scenarios of car following and lane changing, and give warning to users in real time. Users need to know the dynamic environment around them regularly, so that they can take any action in time in case of emergency, so as to avoid long-term maladjustment caused by sudden change of the system.

Psychological comfort
Psychological comfort is to a large extent the subjective feelings of people, the existing theory and technology is difficult to quantify. In this regard, psychological comfort is to study the personalized decision making of IVs, interaction design with other participants and ethical decision making under the condition of solving the reliability and accuracy of the system, so as to improve the trust and acceptance of users.

Security and trust
Research shows that the IVs considering the user's personalized driving style can effectively improve the trust and acceptance of users [69]. According to the driving behavioural such as safe distance, acceleration curve and lane changing speed, drivers' styles can be classified into aggressive type, normal type and cautious type [70]. In order to reduce the tension of users during driving, Lu et al. [71] used the on-board sensing information to learn the driver's speed control experience online and proposed a personalized behavioural learning system (PBLS) to improve the comfort performance of traditional motion planning. The system is based on Neural Reinforcement Learning (NRL) and can adapt to the driving behavioural of different drivers and different driving scenarios. Sama et al. [72] used a DL automatic encoder to process a large number of experienced old drivers' driving data to extract potential features, and then clustered the features into driving behavioural, and created a speed profile to allow IVs to drive according to the user's driving style. Xu et al. [73] proposed a motion planning method for learning natural driving data. On the basis of considering trajectory comfort, efficiency and safety and other factors, combined with lane change decision making of human driving, lane change incentive cost function was established. This method can approach the trajectory of human drivers. Vallon et al. [74] proposed an automatic lane change algorithm based on support vector machine classifier, which can directly capture and copy the natural driving behavioural of human beings, and learn whether to continue to keep the lane or start to change lanes according to the performance preference of drivers. For other road participants, the anthropomorphic driving of IVs can effectively improve the sense of safety. Hang et al. [75] proposed an anthropomorphic decision-making framework FIGURE 7 Improve comfort through external human-computer interaction based on non-cooperative game theory, which not only considers personalized driving style, but also adds social interaction characteristics with other traffic participants, which can cope with complex mixed traffic flow. In addition, external humancomputer interaction (EHMI) enables traffic participants to understand the vehicle's intention more intuitively, so as to avoid psychological panic, as shown in Figure 7. In order to further study the status information requirements of pedestrians for IVs, Faas et al. [76] explored the acceptability of different EHMI variables (No EHMI, State, State + Perception, State + Intention, State + Perception + Intention) to traffic participants. The research results show that state + intention EHMI can increase user experience, perceived intelligence and pedestrian transparency more than other EHMI information. Rettenmaier et al. [77] proposed that in the future mixed traffic environment, when the IVs using EHMI communicates with other traffic participants, the time taken will be significantly shortened and the collision accidents will be less.

Morality and ethics
The development of IVs cannot avoid a series of dilemmas in moral and ethical decision making. For example, in an extreme environment, a smart car needs to decide whether to hurt the lives of users or multiple pedestrians, and any choice will bring about social ethical dilemmas [78]. In addition, IVs will be in the situation of mixed traffic flow for a long time, and will face the problem of responsibility attribution after accidents with other sudden traffic participants due to insufficient perception, violation of traffic rules in order to avoid obstacles etc. In this regard, Waldrop [79] insisted that if there is no clear moral code to guide the decision making of IVs, it is difficult to change the current situation of distrust of users, and at the same time, it will trigger public opinion. In order to quantify moral decision making, Gerdes and Thornton [80] proposed an analogy between the ethical framework of consequentialism and deontology in philosophy and the use of cost function or constraint in optimal control theory. Thornton et al. [81] used a set of moral framework to map the design decision of MPC problem to philosophical principles. By studying how to divide path tracking, vehicle occupant comfort and traffic law priority, the constraints of obstacle avoidance and vehicle turning rate were taken as constraints, which provided guiding principles for the responsibility planning of self-driving vehicles. Riaz et al. [82] put forward a new idea of improving the collision avoidance performance of autonomous vehicles by using human social norms and human emotions, and designed social norms by using emotions as the compliance mechanism, so that IVs can choose the collision with less harm as the decision-making mechanism in possible collisions. In order to obtain data and analysis driving decision-making factors about ethics and law of IVs in moral dilemma, Li et al. [83] combined with the current traffic laws and the cases of accident liability judgment, a series of experiments were carried out in the virtual reality environment. It was concluded that the number of collision targets and whether to comply with the traffic rules were the most important factors affecting the decision making, and they use grey correlation entropy analysis method was used to quantify the severity of collision injury of collision targets. In order to alleviate the severity of inevitable collisions, Wang et al. [84] considered adding potential severity and artificial potential field theory to the controller target to realize IVs emergency obstacle avoidance. On this basis, Wang et al. [85] proposed a Lexicographic Optimization-based model predictive controller, which can avoid obstacles with high assumed priority and solve the problem of moral decision making in vehicle accidents.

ECONOMY OF IVS
Energy saving and emission reduction has always been the focus of the automotive industry, and has become a key step in the large-scale application of IVs. IVs can improve the economy from many aspects, including anthropomorphic driving style changes fuel consumption rate, optimizing driving behavioural through vehicle road collaborative control, cruise control of automatic driving fleet, electrification and sharing of intelligent connected vehicles.

Economical driving behaviour
Driver style has a significant impact on vehicle economy. In order to analyse the relationship between driver behaviour differences and energy consumption in detail, Stogios et al. [86] studied that when driving on highways and under different traffic flows on main roads, fuel consumption on highways can be reduced by 26% when IVs is aggressive driving, while fuel consumption will be increased by 35% under caution driving. Fleming et al. [87] proposed an eco-driving system considering driver's personalized preference. By modelling longitudinal driver's behaviour in the optimal control framework, the balance between driver's preference and energy efficiency goal can be achieved. The control objectives and control protocols of IVs need to be adjusted adaptively according to different driving styles. Lv et al. [88] proposed a collaborative design optimization framework for device parameters and controller parameters of intelligent electric vehicles based on network physical system. At the same time, the driving conditions required by the existing driving style are difficult to meet. Malikopoulas and Aguilar [89] analysed the driving style factors affecting fuel economy and proposed a polynomial meta model framework for optimizing driving style.
With the development of people-vehicle-road collaborative control, the future traffic development trend can be predicted in a short term, such as road shape, traffic flow change, traffic signal status and the movement of adjacent vehicles, so as to optimize driving behavioural and achieve higher fuel economy, as shown in Figure 8. The research shows that, based on the prior knowledge of road speed limit, safe speed on curve and average traffic speed estimation, speed conversion can be more energysaving when the expected speed constraint changes [90]. Therefore, Ding and Jin [91], based on the curvature information of the road ahead of the vehicle extracted from the high-precision digital map, combined with the established fuel consumption model and vehicle dynamics model, applied the dynamic programming algorithm to calculate the optimal speed profile of the entire curved road. And knowing the slope of the road in advance, the vehicle can slide or choose the appropriate speed.
In addition, the impact of eco-driving at signalized intersections on energy efficiency is a hot topic in recent years. It is reported that in 2015, traffic congestion in the United States caused nearly 7 billion hours of delay and more than 3 billion gallons of fuel waste, a large part of which was caused by traffic signal intersection congestion [92]. Hu et al. [93] proposed an optimal vehicle routing algorithm considering waiting time at signalized intersections and ecological driving model. The algorithm is suitable for intersections with dense traffic lights, and the higher the density of traffic lights, the more obvious the advantages of the algorithm. Xu et al. [94] proposed a vehicle speed optimization method based on traffic signal control in intelligent network, which can simultaneously optimize traffic signal timing and vehicle speed trajectory. The method minimizes the total travel time of all vehicles by calculating the optimal traffic signal time and vehicle arrival time, and optimizes the engine power and braking force to minimize the fuel consumption of a single vehicle. The current challenge lies in the local or overall road network traffic signal and vehicle speed trajectory planning. It is necessary to optimize the traffic signal time and the average speed of traffic flow at each intersection to improve the economy and traffic efficiency.

Economical travel mode
The platoon is the prototype of multi-vehicle cooperation. By following each other closely, the air resistance of all vehicles can be reduced, the road traffic capacity can be increased and the fuel economy of vehicles can be improved [95]. Guo and Li [96] proposed a fuel time optimization principle based on Pontriagin's minimum principle to calculate the optimal speed of each vehicle and the speed setting value of the platoon. But the intelligent level of vehicles is low, and information between vehicles cannot be shared, and dense formation is easy to cause traffic accidents. He et al. [97] proposed a multi-stage optimal control scheme considering the length of vehicle queue and the state of traffic lights to obtain the optimal vehicle trajectory on the planned route. Although the queue length and the change of signal state are considered as constraints, it is difficult to estimate the queue length in real time. In this regard, Gao et al. [98] proposed a model based on shock wave perception and back propagation neural network perception, which can predict the queue length of waiting vehicles at signalized intersections in real time under mixed traffic conditions. In order to improve the driving efficiency of IVs in the environment of fuel consumption, Chen et al. [99] proposed a queue path planning strategy based on deep reinforcement learning of network edge nodes, considering the joint optimization problem of task duration and vehicle fuel consumption. Hao et al. [100] proposed a framework combining driving state recognition with queue operation and risk prediction to reduce the interference caused by driving state jitter, so as to improve the evaluation speed, efficiency and fuel economy of multi queue system. IVs is arranged in a long formation to maintain the desired formation while maintaining the safe distance and speed, which requires specific algorithms, controllers and strategies. Soni and Hu [101] summarized a variety of distributed and decentralized vehicle formation control methods, which were divided into leader-follower method, behavioural-based method and virtual structure method. The longitudinal control of vehicle platoon has been studied for many years, and vehicle merging lane changing has always been a hot spot in the research of lateral control. Early merging is based on single platoon MPC, but there are few merging strategies for two-vehicle platoon on adjacent lanes. Min et al. [102] proposed a double platoon merging strategy and designed a DMPC control strategy to control the queue merging problem of expressway. Compared with the traditional single vehicle merging method, the queue merging process is more accurate and time-saving. The non-linear dynamics and safety constraints in vehicle queue are also the research hotspots. He et al. [103] proposed a new distributed economic model predictive control (EMPC) method. By reducing the fuel consumption cost, the strategy ensures the asymptotic stability and leader follower stability of the platoon, and also ensures the fuel economy of the whole platoon.
In addition to the concept of IVs queue, IVs sharing is considered to be the future economic travel mode. Intelligent shared vehicles (ISV) reduce travel time, reduce the cost of passengers per day and per kilometre and reduce carbon emissions to a certain extent [104]. Fagnant and Kockelman [105] proposed an ISV dynamic ride-sharing (DRS) model, which allows two or more users with similar departure, destination and departure time to carpool. The results show that each ISV can replace about 10 traditional family cars, which can effectively reduce the overall vehicle mileage and improve the economy. With the popularization of ISV electrification, the electric vehicles has limited endurance and relatively long charging time. In order to ensure the timeliness of travel, the vehicle scheduling and system rebalancing must consider the tram charging problem. Hu et al. [106] proposed a state of charge (SOC) estimation method of series connected battery pack based on fuzzy adaptive federal filter, which can accurately estimate the remaining power of electric vehicle. Considering the scale and uncertainty of the system prediction, Hu et al. [107] proposes an MPC framework with cost-optimal, which accurately estimates the degradation of fuel cell and battery system. Tang et al. [108] considered the real travel route information data of drivers to train the speed predictor, which can further improve the prediction accuracy, so as to improve the vehicle economy through IVs system control. Ammous et al. [109] modelled the routing problem between multiple charging stations as a multi-server queuing system, and set the goal as a stochastic convex optimization problem, which minimizes the average total travel time of all users relative to their actual travel time. In fact, the specific realization of vehicle to grid (V2G) reduces the energy loss of the whole system by optimizing the charging time and path planning of vehicles. More and more researches are focusing on the coupling characteristics of transportation network and power grid. The optimization problem of single vehicle is changing to the joint optimization of transportation network and power grid, and the performance analysis of charging network based on intelligent sharing platoon has been derived.

CONCLUSION
IVs can make up for the insufficient operation of human drivers, thus reducing traffic accidents, improving road traffic efficiency, providing convenience for vulnerable groups and changing the way of human travel. We review the development of IVs from three important aspects (safety, comfort and economy). On this basis, through a large number of literature survey, we put forward the development status and challenges of IVs in various stages.
First of all, safe driving is the foundation of IVs landing. We outline the framework and information security technology to ensure vehicle safety. Due to the insufficient perception in the existing environment sensing systems, the detection errors and omissions are caused. How to improve the accuracy of the sensing algorithm, the fusion of multi-source heterogeneous sensing data is the primary problem in the research of perception. Moreover, in the face of complex traffic environment and weather changes, the sensor system cannot meet the requirements of over the horizon sensing. In the foreseeable future, the cooperative sensing of communication technology and sensor system will effectively improve the accuracy of sensing data. At the same time, the decision-making system is faced with the uncertainty of the rules of the unknown risk scenarios, so it is urgent to study a reliable risk assessment and prediction model, and the pedestrian trajectory prediction is the difficulty. There are some problems such as the deviation between the control target of the decision system output and the ideal goal of the control execution system, and the function limitation of the actuator. How to consider various vehicle dynamic instability constraints to establish control model and actuator failure redundant mechanism design are challenging problems. In addition, although the IVs networking can make up for the sensor perception defects and improve the accuracy of decision making and control, the information security problems cannot be ignored. How to build multi-domain layered intrusion detection system and active protection information security model and 'end management cloud' information security protection system become the next research focus.
Secondly, from the perspective of human factors and social acceptance, the status quo and challenges of IVs comfort are discussed. The influencing factors of users' physiological comfort mainly come from the mechanization of decision control algorithm, the great difference of speed change and the impact of unknown scene. One of the biggest challenges for IVs is not only to drive safely, but also to drive as smoothly as an old driver. Therefore, it is necessary to use DL algorithm to refer to human driving experience for path planning and trajectory tracking, so as to improve the driving proficiency of IVs. At the same time, it is necessary to consider the speed control of physiological tolerance of most people. However, it is difficult to fully consider people's psychological feelings. Therefore, learning the driving style accepted by users or most users and the EHCI that is easier to be understood by other traffic participants can effectively improve the degree of trust; in the moral and ethical dilemma, to develop a complete set of decision-making standards to meet the acceptable level of users and society, and improve the fault tolerance rate of users are the key research objects in the future.
Finally, for the economy of IVs, we describe the economic driving behaviour and travel mode from the micro and macro level. Although driving style can improve the economy, it is still difficult to balance the driving conditions, energy efficiency and personalized preference. In this regard, the use of peoplevehicle-road integrated collaborative control can deal with different road scenarios and optimize driving behaviour. For the IVs queue, the longitudinal and horizontal coupling control, multi-queue cooperative control, queue stability control and other aspects need to be studied in depth. With the sharing and electrification of IVs, it is necessary to solve the vehicle sharing mechanism, maximize the traffic efficiency and the coupling characteristics of transportation network and power grid brought by V2G technology. Especially in the next stage, the IVs are the coordinated control of vehicle, road, network and cloud. Accelerating the exploration of the integration of smart city, smart transportation and smart vehicle (SCSTSV) can improve the economy of the whole transportation system. This paper systematically combs the key technologies in all aspects of the development of IVs, hoping to promote the rapid development of IVs and provide systematic understanding for researchers in various disciplines.