1. Introduction
Software-defined vehicle (SDV) is the general trend of the automotive industry, with the concept of decoupling software and hardware and creating a new research topic of “atomic” technology combination [
1]. In the past, the architectures of autonomous vehicles have realized one function with a set of parts and components. The typical feature was that hardware and software were coupled. With such an architecture, automakers purchase software functions interfaced with hardware, the corresponding hardware cannot be changed. Meanwhile, the hardware cannot be reused, and a specific function is not available without a combination of specific hardware and software, and therefore, the cost is relatively higher. Moreover, there is no space or possibility for vehicles to evolve. New functions and services require the addition of corresponding parts and software. Therefore, future architectures of autonomous vehicles require the decoupling of software and hardware. Hardware would be translated into atoms, which would be orchestrated to realize safety functions. This so-called hardware atomization means translating hardware that has been decoupled with software into technological “atoms” to be revoked by various softwares [
1,
2,
3]. Autonomous vehicles with an architecture in which software and hardware are decoupled would have the following advantages [
4]: First, original equipment manufacturers (OEMs) could flexibly replace hardware, since software and hardware are decoupled. Second, hardware could be reused. After the standardization and abstraction of hardware, one hardware could realize several functions, thus, reducing the number of hardware units and saving cost. Third, OEMs could add software to add more functions, without the need of replacing or adding hardware, leaving more space and possibilities for vehicles to evolve. The so-called atomization technologies refer to minimal hardware units of autonomous vehicles, such as LiDAR, cameras, the steering system, and the braking system. Such units are combined and orchestrated to realize the functions of autonomous vehicles [
1]. Since the safety and cost of autonomous vehicles would vary based on the “atomic” technology combination, there is a need to study the safety and cost of autonomous vehicles brought by the combination of “atomic” technologies.
Vehicle to everything (V2X) has gradually become an industry-wide consensus, especially in China [
5,
6], and has generated a new research topic of how to distribute capabilities and costs taking into consideration vehicles and roads. From the perspective of technology, autonomous vehicle technologies are complex and there will always be long-tailed problems. However, V2X can improve the intelligent capabilities of vehicles and provide redundancy. From the perspective of cost, a higher level of intelligence would increase the cost of vehicles, making it hard to promote the uptake of autonomous vehicles, which could fall into the “nobility” vehicles. V2X could reduce the cost of vehicles and could turn autonomous vehicles into a reality by migrating relevant hardware such as cameras, MMW radars, and LiDAR to roads to provide intelligent functions for vehicles [
7]. Therefore, the cost of such hardware would be shared by thousands of vehicles and the utilization would also increase. Otherwise, the cost would be woven into the total cost of autonomous vehicles. From the perspective of benefits, autonomous vehicles can only improve themselves and generate local benefits, failing to improve traffic flow. However, V2X would make global improvements and would maximize social benefits. The essence of V2X is to distribute “intelligent capabilities and costs” to vehicles and roads, which requires the analysis of the safety and cost under different combinations of “atomic” technologies.
The safety mechanism of autonomous vehicles is to use different safety functions to avoid accidents and to improve the safety of vehicles. There has been a lot of research on the safety effectiveness of the safety functions of autonomous vehicles [
8]. Regarding automatic emergency braking (AEB), most studies have only focused on rear-end crashes. Some studies have mainly focused on pedestrian crashes and cyclist crashes. According to the available evidence, the effectiveness of AEB in avoiding target crashes ranges from 18% to 72% [
9,
10]. For adaptive cruise control (ACC), the effectiveness of ACC in avoiding a rear-end crash and other related crashes ranges from 12% to 16%, which is a relatively low level [
11,
12]. For lane keeping assist (LKA) and lane departure warning (LDW), lane departure-related crashes could be avoided by using LKA and LDW, including single crashes, front crashes, sideswipe same direction crashes, and sideswipe opposite direction crashes. The effectiveness of LDW is in the range of 10–48% [
13]. The effectiveness of LKA is in the range of 20–51% [
14]. For blind spot detection (BSD), according to the related references, the target crash scenario that could be avoided by BSD is the lane-change crash and the crash avoidance effectiveness of BSD ranges between 14% and 58% [
15,
16]. For connected intersection movement assist (IMA), the intersection crashes that could be avoided by IMA, include straight crossing paths at non-signal (SCP), left turn into the path at non-signal (LTIP), right turn into the path at signal (RTIP), running a red light, and running a stop sign. The crash avoidance effectiveness of IMA is in the range of 23–67% [
17,
18]. For connected left-turn assist (LTA), the left turn across the path crash could be avoided by using LTA technology. The effectiveness of LTA is in the range of 32% to 60% [
19,
20]. It should be noted that using V2X to realize connected AEB (C-AEB), connected LKA (C-LKA), connected ACC (C-ACC), and connected LDW (C-LDW) would result in better crash avoidance effectiveness [
9,
21]. In addition to avoiding collisions, mitigating harm of collision is also a research direction. Parseh and Asplund proposed a collision reconfiguration system that could mitigate harm due to collision by changing where vehicles were hit and how the impact force was directed towards the vehicle [
22]. Test scenarios for autonomous vehicles have received a lot of attention. A systematic and data mining approach was developed to extract and generate high-risk precrash test scenarios from the data [
23]. Speed and space perception also affects the occurrence of road traffic accidents [
24]. Yu developed a binary logit regression model to differentiate the hazardous scenarios of autonomous vehicles [
25]. The model was developed to quickly generate various test scenarios by adjusting various traffic scenarios and different severities of quantifiable weather and interference parameters [
26].
Under the new development trends of V2X and SDV, the safety effectiveness of autonomous vehicles involves numerous “atomic” technology combinations and the coupling of multiple safety functions, therefore, requiring a method to evaluate the safety effectiveness based on the coupling of multiple variables. As compared with previous research, to evaluate the safety effectiveness of autonomous vehicles, studies should go to a deeper level, namely, evaluate the safety effectiveness of “atomic” technologies including vehicle-side cameras, millimeter-wave (MMW) radar, LiDAR, high-precision positioning units, automotive computing units, the brake-by-wire system, the steer-by-wire system, the on-board unit (OBU) for communication, as well as roadside cameras, MMW radars, LiDAR, roadside units (RSUs) for communication, and computing units [
1]. Among these “atomic” technologies, “atomic” sensing technologies are the foundation and the core of securing the safety of autonomous vehicles, and are selected as the research object in this paper. The “atomic” technologies in this paper mainly focus on the sensing system, including cameras, MMW radars, as well as vehicle-side and roadside LiDAR. Only by atomizing technologies can they be recombined. The “atomic” technology combinations mean orchestrating them. Safety effectiveness is reflected by the crash avoidance effectiveness of vehicles. Different types of “atomic” technologies in different positions can enable various types of safety functions of autonomous vehicles, which can avoid a certain proportion of target accidents, thus, generating the comprehensive crash avoidance effectiveness of vehicles. In this paper, we evaluate the safety effectiveness and cost of autonomous vehicles from such a perspective.
In this paper, we provide an “atomic” technology safety effectiveness evaluation model with the coupling of multiple variables based on “atomic” technologies, safety functions, target accident types, and crash avoidance effectiveness, and we provide an “atomic” technology cost-sharing model based on the traveled distance during the life cycle of vehicles and based on the traffic flow over the life cycle of roads. We apply the models to mainly answer the following questions in a systematic manner:
- (1)
How can the crash avoidance effectiveness of different “atomic” sensing technologies be quantified realizing various safety functions?
- (2)
What is the crash avoidance effectiveness of the “atomic” sensing technology combinations? What would the life-cycle cost be of the combinations?
- (3)
What is the “atomic” technology combination to meet the safety requirements at the lowest cost? What is the “atomic” technology combination to feature the highest crash avoidance effectiveness with a certain cost.
Some summarized highlights of the results include: (1) Roadside sensors can result in a higher comprehensive traffic crash avoidance effectiveness and a lower cost to use per kilometer than vehicle-side sensors. (2) The cost would increase with the addition of “atomic” sensing technologies on the vehicle side, while the increase in crash avoidance effectiveness would slow down. (3) It is necessary to switch to V2X and introduce roadside “atomic” technology combinations to realize better safety effectiveness at a lower cost for vehicles. (4) From the perspective of crash effectiveness and lift-cycle cost sharing, the “atomic” technology combinations for V2X would be superior to the “atomic” technology combinations only for vehicles.
2. Methodology
To measure the safety effectiveness of autonomous vehicles enabled by the combination of different “atomic” technologies, in this paper, we provide a safety effectiveness evaluation model based on multiple variables of “atomic” technologies, safety functions, accident types, and crash avoidance effectiveness, as shown in
Figure 1 and Equations (1)–(5). There are three steps. The first step is to quantify the safety effectiveness and target accident types of each “atomic” technology. It is necessary to identify multiple safety functions that can be realized by the “atomic” technology. Then, we calculate corresponding crash avoidance effectiveness for the target accident types of the “atomic” technology, realizing several safety functions, as shown in Equation (5). The second step is to quantify the target type of accidents and corresponding coupled crash avoidance effectiveness of the “atomic” technology combinations. Each “atomic” technology in the combination has target accident types and corresponding avoidance effectiveness. For the accident type that is listed as the target by multiple atomic technologies, the comprehensive collision avoidance effectiveness can be calculated by using Equation (4). The third step is to weigh the type of accident and the crash avoidance rate of the “atomic” technology combination according to the proportions of different types of accidents in all traffic accidents in China and obtain the comprehensive crash avoidance rate, as shown in Equation (3). As shown in Equation (1), the expression of “atomic” technology combinations encompasses 14 types of “atomic” technologies. Equation (2) is the calculation method for the unit cost shared by “atomic” technology combinations, with the key in the cost of every “atomic” technology and the number of technologies. Next, the details of the data are introduced.
More details of the model are shown in
Figure 2. The automotive “atomic” technologies include front-facing cameras, front-facing MMW radars, front-facing LiDAR, rear cameras, rear MMW radars, rear LiDAR, side cameras, side MMW radars, side LiDAR, top LiDAR, OBUs, roadside cameras, roadside MMW radars, and roadside LiDAR. All these “atomic” technologies have factored into the applicability under different weather and light conditions. In this paper, “atomic” technologies only focus on the key sensing hardware of autonomous vehicles, and future research will focus on computing chips, steer-by-wire systems, and brake-by-wire systems. There are 26 automotive safety functions, such as front collision warning (FCW), AEB, front cross traffic brake (FCTB), rear collision warning (RCW), rear cross traffic brake (RCTB), ACC, LKA, LDW, BSD, lane change assist (LCA), advanced LKA (A-LKA), and advanced IMA (A-IMA), and there are 26 V2X-based safety functions, such as connected FCW (C-FCW), C-AEB, connected FCTB C-FCTB), C-ACC, C-LKA, connected LCA (C-LCA), connected and advanced LKA (CA-LKA), and connected and advanced IMA (CA-IMA). Every safety function has a corresponding target accident type and crash avoidance rate, for which a database has been created in the previous papers. According to the report on road traffic accidents in China, there are 14 types of accidents, such as head-on collisions, rear-end collisions, side collisions, scrapping, and pedestrian/cyclist collisions. The output crash avoidance effectiveness is divided into the comprehensive crash avoidance effectiveness of autonomous vehicles and of V2X, which can reflect the safety effectiveness of vehicles and roads. The underlying safety-related data are from the safety functions’ crash avoidance effectiveness of nearly one hundred research papers on safety functions and the proportions of accidents listed in the official reports.
where:
SensorGroup is the combination of “atomic” technologies, which is made up of 14 variables (xj), x1, x2, …, x13 represent the number of front cameras, front MMW radars, front LiDAR, rear cameras, rear MMW radars, rear LiDAR, left-right cameras, left-right MMW radars, left-right LiDAR, top LiDAR, OBU, roadside cameras, roadside MMW radars, and roadside LiDAR, respectively (If x1 = 2, it means there are two front cameras and if x7 = 1, it means using four cameras to cover the field of view on the right and left);
Costsensorgroup is the cost per kilometer shared by sensorGroup;
UCvehiclesensor,j is the cost per kilometer shared by vehicle-side “atomic” technologies; UCRoadsensor,j is the cost per kilometer shared by roadside “atomic” technologies.
CCAESensorgroup represents the comprehensive collision avoidance effectiveness of sensorGroup;
CAEk,Sensorgroup represents the avoidance effectiveness of the accident (k);
CPk is the proportion of the type of accident (k) in all traffic accidents in China;
CAEk,j is the collision avoidance effectiveness of the “atomic technology” (j) to the type of accident (k);
SFj,i represents whether the “atomic” technology (j) can realize the safety function (i) (1 means positive and 0 means negative);
Ei is the collision avoidance effectiveness of the safety function (i);
FCi,k represents whether the type of accident (k) is the target accident type of the safety function (i) (1 means positive and 0 means negative);
Wj represents whether the “atomic” technology (j) can work in inclement weather (1 means positive and 0 means negative);
Lj represents whether the “atomic” technology (j) can work in darkness (1 means positive and 0 means negative);
CWPw is the proportion of traffic accidents in China under inclement weather;
CLPl means the proportion of traffic accidents in China in darkness;
K represents the type of accident, and k = 1, 2, 3, …, 14 means frontal crash, rear-end crash, left-turn crash at a crossing, right-turn crash at a crossing, straight running crash at a crossing, off road obstacle crash, side crash not at a crossing, scrapping, stationary vehicle crash, other vehicle-vehicle crash, pedestrian/cyclist collision, on-road obstacle crash, off road obstacle crash, rollover/rolling/crash, and other vehicle accidents, respectively;
i represents 52 safety functions, including AEB, ACC, LKA, C-AEB, C-ACC, and C-LKA.
Figure 2.
Details of the multivariable coupling model: “Atomic” technology, safety function, accident type, and crash avoidance effectiveness.
Figure 2.
Details of the multivariable coupling model: “Atomic” technology, safety function, accident type, and crash avoidance effectiveness.
5. Discussion
In this paper, we developed an “atomic” technology safety effectiveness evaluation model with the coupling of multiple variables based on “atomic” technologies, safety functions, target accident types, and crash avoidance effectiveness, we evaluated the comprehensive collision avoidance effectiveness of cameras, millimeter-wave radars, and LiDARs placed on the front, rear, side, and top of the vehicle, and cameras, millimeter-wave radars, and LiDARs placed on the roadside, and we also evaluated the comprehensive collision avoidance effectiveness corresponding to different “atomic” technology combinations. In total, 14 types of sensors, 52 safety functions, 14 types of accidents, and the applicability of sensors to bad weather and bad light were considered in the model. In terms of cost, in view of the difficulty in quantifying the cost of vehicle-side and roadside “atomic” technologies, we developed a cost-sharing model based on the traveled distance during the life cycle of vehicles and based on the traffic flow over the life cycle of roads to evaluate the unit cost per km of different “atomic” technology combinations. We quantified the comprehensive collision avoidance effectiveness and the cost of typical vehicle-side and roadside “atomic” technology combinations, drew a panoramic picture of the comprehensive collision avoidance effectiveness and the unit cost per km of all “atomic” technology combinations, and selected the optimal “atomic” technology combination with the lowest cost under given safety requirements and the optimal “atomic” technology combination with the highest safety effect under given cost requirements.
The results clearly show that from the perspective of improving safety effectiveness and reducing usage costs, the V2X technology path is the first choice. The autonomous vehicle path has limitations. As the cost rises, the room for improving safety effectiveness becomes smaller. However, V2X can significantly improve the comprehensive collision avoidance effectiveness at a lower cost. At the same time, due to the existence of roadside “atomic” technologies, in order to achieve the same safety effectiveness, vehicle-side “atomic” technologies can be significantly reduced. Increasing roadside configurations and reducing vehicle-side configurations will become an important and promising direction for the development of autonomous vehicles. V2X has cost and safety advantages, but since autonomous vehicles need to drive on autonomous roads to obtain the benefits of V2X, it is necessary to retain sufficient vehicle-side “atomic” technology combinations to achieve sufficient accident avoidance effectiveness, especially in the initial stage of autonomous road construction.
The problem of selecting optimal “atomic” technology combinations was discussed in this paper. Under the given safety effectiveness, the life-cycle cost of “atomic” technology combinations based on V2X was much lower than that of “atomic” technology combinations based on autonomous vehicles. In this paper, we proposed the optimal “atomic” technology combination with the lowest cost under the assumption of 80% avoidance effectiveness. Under the given cost limit, “atomic” technology combinations with higher roadside configurations can achieve a higher comprehensive collision avoidance effectiveness than “atomic” technology combinations based on autonomous vehicles. In this paper, we also proposed the optimal “atomic” technology combination with the highest safety effectiveness under the constraint of an RMB 5000 life-cycle cost. From the perspective of safety effectiveness and life-cycle shared cost, the results of this paper indicate that V2X will become the future development direction.
This study also points out the directions of future research that need to be studied urgently for promoting the development of the V2X. First, research is needed to quantify the safety impact of penetration rate of autonomous vehicles with different “atomic” technology combinations and the coverage rate of autonomous roads with different “atomic” technology combinations on the annual traffic accident casualty at the country level. In the future, fleets will be composed of traditional vehicles and autonomous vehicles, and roads will consist of traditional roads and autonomous roads. In this paper, we focused on “atomic” technologies. However, the higher-level topic is about the impact of autonomous vehicle deployment and autonomous road deployment on the country’s annual casualties caused by traffic accidents, which would require us to calculate the corresponding economic benefits and total costs from the national level to support further decisions on technology paths. Second, future research should study and formulate corresponding autonomous road policies from the perspective of the government. China has introduced many policies to guide the construction of autonomous roads. Beijing and many other cities have designated special road areas to build autonomous roads in a bid to promote the further development of V2X-based autonomous vehicles. Third, in terms of technology R&D, it is necessary to focus on how to rely on the information of roadside “atomic” technologies to enable the safety functions of autonomous vehicles. In this case, we need to consider time delay, information fusion, reliability, and other scientific issues. Fourth, we need to study the V2X-based business model. There are 5.28 million kilometers of various roads in China. The construction of autonomous roads requires huge costs. Who should pay for the construction costs? How should we recover the costs and achieve profitability in the future?
The limitations of this paper are as follows: First, computing chips, steer-by-wire, brake-by-wire, and some other safety-related basic hardware of autonomous vehicles are not considered in the model, which are also important for supporting safety functions. In fact, the computing power required by chips depends on the size of the information that the vehicle needs to process. When vehicle-side sensors are piled up, chips with large computing power become necessary. Second, in this paper, we fail to consider the impact of more detailed parameters of cameras, millimeter-wave radars, and LiDARs on safety effectiveness. The main modeling logic in this paper is that vehicle-side and roadside “atomic” technologies can support various safety functions, and these safety functions can avoid a certain proportion of target accident types. Due to the lack of data and relevant modeling methods, it is difficult to consider the detailed parameters of various sensors in the research. With more data to be obtained, the model provided in this paper can be further expanded. Third, the effectiveness of the safety functions could be updated if more research on the crash avoidance effectiveness of these safety functions are published. The effectiveness of the same safety function reported by different studies is quite different because studies have used different methods, data and detailed condition parameters. Therefore, the more research results that are considered in a meta-analysis, the more accurately the effectiveness of safety functions can be calculated.