Cooperative Environmental Perception Task Offloading for Connected and Autonomous Vehicles

Abstract

Cooperative environmental perception is an effective way to provide connected and autonomous vehicles (CAVs) with the necessary environmental information. The research goal of this paper is to achieve efficient sharing of cooperative environmental perception information. Hence, a novel vehicular edge computing scheme is proposed. In this scheme, the environmental perception tasks are selected to be offloaded based on their shareability, and the edge server directly delivers the task results to the CAVs who need the perception information. The experimental results show that the proposed task offloading scheme can decrease the perception information delivery latency up to 20%. Therefore, it is an effective way to improve cooperative environmental perception efficiency by taking the shareability of the perception information into consideration.

1. Introduction

In recent years, connected and autonomous vehicles (CAVs) have become a hot research topic in both academia and in the industry [1,2,3,4]. For CAVs, environmental perception is a fundamental requirement. Although current CAVs have been equipped with various sensors (e.g., cameras, radars, and lidars) for environmental perception [5,6,7], there still exist blind areas that CAVs cannot perceive in certain scenarios (e.g., obstacles, bend in a road, etc.). Therefore, cooperative environmental perception [8] is an effective way to enhance the perceptive capabilities of CAVs. With cooperative environmental perception, a CAV can obtain the environmental information shared by other CAVs even if the CAV cannot perceive the environmental information directly.

Since the timeliness of the environmental perception information is vital for driving safety, achieving efficient cooperative environmental perception is the research goal of this paper. Cooperative environmental perception can be divided into the following two parts.

(1) Sensing data processing. The sensors on a CAV may generate 2 GB of real-time data per second on average [9]. Hence, it is difficult to guarantee that all the sensing data can be processed by the CAV within the tolerable delay [10]. To overcome this limitation, vehicular edge computing (VEC) [11] has become a promising computing paradigm. Resource-constrained CAVs can offload some perception tasks, especially those that are delay-sensitive and computing power-consuming, to edge servers for computation.

(2) Environmental information delivery. After the sensing data are processed, the environmental information extracted from the sensing data is sent to the CAVs that need the information to make driving decisions.

In traditional schemes [12,13,14,15,16,17,18,19], once a CAV offloads a task to the edge server, the computation result of the task is returned to the CAV. After that, the CAV transmits the result to the specific CAVs or simply broadcasts the result to its adjacent CAVs, which is inefficient.

In this paper, the shareable characteristics of the environmental perception task results are considered, and a novel vehicular edge computing scheme is established. According to this scheme, if the result of an environmental perception task is required by more CAVs in the surrounding areas, the task is offloaded to the edge server with a higher priority. Moreover, the edge server delivers the task result directly to the CAVs that require it, which reduces delivery latency. We first design an offloading strategy based on the greedy algorithm, which offloads tasks according to the shareable level of environmental perception tasks. Since this strategy does not consider the load of the vehicle, an improved version of the task offloading algorithm considering load balance is also proposed. The simulation results show that the task offloading strategy proposed in this paper makes full use of the shareable features of the environmental perception task and effectively improves the timeliness of environmental perception information.

The main contributions of this paper are summarized as follows:

(1) The concept of environmental perception task shareability for CAVs is created, and the shareability coefficient is designed to quantify the shareability of environmental perception tasks.

(2) Aiming at the straight road and the intersection scenario, respectively, the requiring area model for the environmental perception task of CAVs is established.

(3) An edge computing architecture is proposed for the shareable perception tasks of CAVs, in which the computational results of the tasks are directly delivered to the requiring CAVs to effectively reduce transmission latency.

(4) Two environmental perception task offloading strategies are designed, including the greedy offloading strategy and the load balancing offloading strategy. The effectiveness and efficiency of the two strategies are verified through extensive simulation experiments.

To the best of our knowledge, this is the first attempt to consider the shareability of the task result in a vehicular edge computing context.

The remainder of the paper is organized as follows. Related works are summarized in Section 2. The system model is described in Section 3. The designed algorithms are presented in Section 4 in detail. The experimental results are shown and analyzed in Section 5. Finally, we conclude the paper in Section 6.

2. Literature Reviews

2.1. Cooperative Perception for CAVs

Cooperative perception technology is very helpful for CAVs to promote the environmental perception ability. In Reference [20], a comprehensive survey of state estimation and motion prediction schemes for vehicles and pedestrians is provided. The survey takes the cooperative perception of CAVs as an application scenario and puts forward the key challenges faced by the cooperative perception of CAVs. Yuan et al. study the sharing of cooperative perception data between vehicles to improve perception accuracy and reduce the communication overhead of CAVs [21]. Lei et al. propose the first latency-aware cooperative perception system from the perspective of machine learning in [22], which enhances the robustness and effectiveness of cooperative perception. To overcome the shortage of irrelevant information in the cooperative perception scheme between vehicles, a cooperative perception system called augmented informative cooperative perception is proposed in [23], which can filter irrelative information quickly and optimize the informativeness of shared data between vehicles to improve the fused presentation.

The above research works explore the cooperative perception of CAVs, but edge computing architecture is not considered in these works. In the paper, we combine cooperative perception technology and edge computing to make full use of their advantages, thus further promoting the safety and task-processing ability of CAVs.

2.2. Perception Task Offloading for CAVs

When CAVs are not able to process environmental perception tasks in a timely manner, they can offload the tasks to edge servers. Ahmed et al. make a comprehensive survey on task offloading in vehicular edge computing environments [24]. In [25], each offloaded perception task is assigned a priority, so that the edge server can process more tasks with high priorities within a certain time. To address the safety threats arising from obstructed lines of sight, Xiao et al. propose a multi-tier perception task offloading framework with a collaborative computing approach in [26]. In this way, a CAV can achieve comprehensive perception of the concerned region-of-interest by leveraging collaborative computation with nearby CAVs and road side units. Krijestorac et al. design a hybrid offloading strategy mixing up the horizontal offloading (vehicle-to-vehicle) and vertical offloading (vehicle-to-edge server) to promote the speed of the task processing [27]. In order to solve the problem of overloaded edge servers caused by the increasing number of vehicles, Feng et al. design a reverse offloading scheme [28]. After receiving the environmental perception information uploaded by the vehicles, the edge servers reverse offload the computing task to vehicles. In this way, the computing loads of the edge servers are reduced. Jin et al. present a context-aware offloading strategy based on a differential evolution algorithm [29], which considers vehicle mobility, roadside unit (RSU) coverage, vehicle priority, etc. They solve the joint optimization problems of the offloading decision and resource allocation. Sun et al. consider collaborative task offloading between vehicles in [30], and present an offloading algorithm based on adaptive learning. The algorithm enables vehicles to learn the offloading latency performance of their neighboring vehicles, and the goal is to minimize the average offloading delay. The fact is considered in [31] that the vehicles travels a distance during the time period between offloading sensing data and receiving the computation results. The offloading order of the tasks is determined by vehicle mobility and the computation capacity of the edge server, as well as the aforementioned distance effectively. Qin et al. propose an approach to jointly schedule task offloading and result caching to minimize data duplication [32]. This research considers the reusability of perception task results; however, it ignores the dynamic features of the sensing data and does not make sure that the results are delivered accurately and in a timely manner to vehicles which need the results.

The research works mentioned above do not focus on the shareability of the results of the environmental perception task, i.e., there may be multiple CAVs in the scenario that also need the results of the environmental perception task. In this paper, we analyze the shareability of the environmental perception results and design an offloading strategy which is significantly different from the existing studies.

3. System Model

3.1. System Architecture

A typical application scenario considered in this paper is shown in Figure 1. An edge server and multiple access points (APs) are deployed at the roadside, and the APs are connected to the edge server. The APs provide wireless access service for the CAVs, and the edge server can execute environmental perception tasks offloaded by the CAVs. A CAV can offload its environmental perception tasks to the edge server through an AP and can obtain the perception results from another appropriate AP even if it leaves the original AP.

In traditional vehicular edge computing solutions for cooperative perception, the task results are returned to the vehicle that generates the task, and the vehicle then delivers the results to other vehicles who demand the results through V2V communications, as shown in Figure 2a. Different from the traditional scheme, in our design, the edge server directly delivers the task results to the vehicles with demands, as demonstrated in Figure 2b. With such a design, the latency of the task result delivery is greatly reduced.

3.2. Perception Task Model

Assume that there are n CAVs in the considered scenario, denoted by a set $V=\left\{v_1,v_2,\dots ,v_n\right\}$. Each CAV needs to perceive environmental information from four directions, i.e., front-side, rear-side, left-side, and right-side. Hence, the CAV $v_i(i=1,2,\dots ,n)$ generates four perception tasks in a perceptive cycle, denoted by a set ${\mathrm{T}}_i=‘\left\{{\tau }_{\mathrm{front}}\left(i\right),{\tau }_{\mathrm{rear}}\left(i\right),{\tau }_{\mathrm{left}}\left(i\right),{\tau }_{\mathrm{right}}\left(i\right)\right\}$ where $i=1,2,\cdots ,\mathrm{n}$, as shown in Figure 3. Therefore, the set of tasks generated by all CAVs in this scenario in a perceptive cycle is ${\mathrm{T}}_{\mathrm{all}}=\cup {\mathrm{T}}_i(i=1,2,\dots ,n)$. The maximum tolerable latency of each perception task $\tau$ is denoted as $\Delta$.

3.3. Requiring Areas of Perception Tasks

In this paper, a new concept is proposed, which is requiring area. For a perception task, its result is needed by the CAVs in its requiring areas. The requiring areas of perception tasks are analyzed using two common traffic scenarios, i.e., straight roads and intersections. As indicated in Figure 4, the shaded regions in the four subplots represent the requiring areas for the perception tasks ${\mathrm{T}}_i=\left\{{\tau }_{\mathrm{front}}\left(i\right),{\tau }_{\mathrm{rear}}\left(i\right),{\tau }_{\mathrm{left}}\left(i\right),{\tau }_{\mathrm{right}}\left(i\right)\right\}$ generated by the task vehicle from the four directions in the straight road scenario. Taking Figure 4a as an example, the environmental information directly in front of the task vehicle (i.e., the vehicles parking on the side of the road, the pedestrians crossing the road, etc.) has an impact on the vehicles driving in the same direction behind the task vehicle and also on the oncoming vehicles towards the lane. Therefore, the vehicles in the shaded area in Figure 4a need the front-side environmental perception information of the task vehicle.

Similarly, the shaded areas in the four subplots in Figure 5 represent the requiring areas for the perception tasks ${\mathrm{T}}_i=\left\{{\tau }_{\mathrm{front}}\left(i\right),{\tau }_{\mathrm{rear}}\left(i\right),{\tau }_{\mathrm{left}}\left(i\right),{\tau }_{\mathrm{right}}\left(i\right)\right\}$ generated by the task vehicle from the four directions in the intersection scenario. In Figure 5a, for example, the front-side environmental information of the task vehicle has an impact on vehicles that are about to enter the intersection from all directions, while it does not affect the vehicles that have already left the intersection. Therefore, the four shaded areas in Figure 5a are the requiring areas for the front-side environmental perception task ${\tau }_{\mathrm{front}}$ of the task vehicle.

The environmental perception tasks $\tau$ generated by a task vehicle are assigned different weights $w_{\tau }$ according to the number of vehicles in the requiring areas. The more vehicles exist in the requiring areas of a task, the higher the weight is that is assigned to the task. The weight of task $\tau$ is defined as the ratio of the number of vehicles in the requiring area $k_{\tau }$, and the total number of vehicles n in the scenario is defined as

$$w_{\tau }=\frac{k_{\tau }}{n}$$

(1)

3.4. Shareability Factor of the Task Result

Since the computation resource on the edge server is also limited, it is not practical to offload all tasks to the edge server when there are too many vehicles. Therefore, only a part of the tasks can be offloaded. The set of offloaded tasks is denoted as $\mathrm{O}$, where $\mathrm{O}\subseteq {\mathrm{T}}_{\mathrm{all}}$, and the number of offloaded tasks is $\left|\mathrm{O}\right|$.

Another new concept that is part of the shareability factor is also created to measure the level of sharing for each environmental perception task $\tau$, which is denoted as $\theta$. The physical meaning of the shareability factor $\theta$ is that an offloaded task result is needed by $\theta$ vehicles on average. Hence, a larger $\theta$ means a higher efficiency for the offloaded task selection.

Since the number of vehicles in the requiring areas of task $\tau$ is $k_{\tau }$, the shareability factor is computed as

$$\theta =\left\{\begin{array}{cc}\frac{{\sum }_{\tau \in \mathrm{O}}{k}_{\tau }}{\left|\mathrm{O}\right|}& \mathrm{O}\ne \varnothing \mid \\ 0& \mathrm{O}=\varnothing \end{array}\right.$$

(2)

where ${\sum }_{\tau \in \mathrm{O}}{k}_{\tau }$ represents the sum of the number of vehicles in the requiring areas of each task that is offloaded.

3.5. Delay Model

Since efficiency is significant for environmental information perception, the offloading delay from the generation of a task to the delivery of its result is analyzed. The delay includes two parts, i.e., the transmission delay and the processing delay.

3.5.1. Transmission Delay

In the scenario considered in this paper, the APs are installed along a road to provide seamless coverage, and multiple APs connect to an edge server. The CAVs transmit the tasks to the edge server or receive the results from the edge server through the APs. It is assumed that x tasks are needed to be uploaded to the AP within an AP communication range, but only y subchannels are available for the transmissions ($y<x$). Thus, the $M/M/y$ queuing model is adopted to express the process of task uploading. Assume that the data arrival rate is $\lambda$, and its service time is exponentially distributed with a mean value of $\frac{1}{\mu }$. Thus, the waiting time of task transmission can be denoted as

$$t_{w_{up}}\left(\tau \right)=\frac{\rho \varsigma }{(1-\rho )\lambda }$$

(3)

where $\rho$ is the service utilization factor and $\varsigma$ denotes the queuing probability. In the $M/M/y$ queuing model, the following equations holds:

$$\rho =\frac{\lambda }{y\mu }$$

(4)

$$\zeta =\frac{{\left(y\rho \right)}^y}{y!(1-\rho )}{\left[1+\frac{{\left(y\rho \right)}^y}{y!(1-\rho )}+\sum _{i=1}^{y-1}\frac{{\left(y\rho \right)}^i}{i!}\right]}^{-1}$$

(5)

The uplink transmission rate $r_{\mathrm{up}}$ and the downlink transmission rate $r_{\mathrm{down}}$ can be derived from Shannon’s formula as

$$r_{\mathrm{up}}=B_{\mathrm{up}}{log}_2\left(1+\frac{P_v{d}^{-\alpha }{h}_{\mathrm{up}}}{N}\right)$$

(6)

$$r_{\mathrm{down}}=B_{\mathrm{down}}{log}_2\left(1+\frac{P_a{d}^{-\alpha }{h}_{\mathrm{down}}}{N}\right)$$

(7)

where $B_{\mathrm{up}}$ and $B_{\mathrm{down}}$ denote the uplink and downlink transmission bandwidths, respectively; $P_v$ and $P_a$ denote the transmission power of the vehicle and AP, respectively; d denotes the distance between the vehicle and AP; $\alpha$ denotes the path loss index; the channel gains of the uplink and downlink are $h_{\mathrm{up}}$, $h_{\mathrm{down}}$, respectively; and N is the environmental noise of the system. Therefore, the uplink transmission delay and downlink transmission delay are expressed as

$$t_{\mathrm{up}}\left(\tau \right)=\frac{l_{\tau }}{r_{\mathrm{up}}}$$

(8)

$$t_{\mathrm{down}}\left(\tau \right)=\frac{d_{\tau }}{r_{\mathrm{down}}}$$

(9)

where $l_{\tau }$ is the size of the perception task $\tau$ (in bits), and $d_{\tau }$ is the size of the task result.

It is assumed that the edge server has an appropriate mobility management scheme. Hence, when the edge server finishes the task processing, it can deliver the task result to the APs that cover the requiring areas of the task. These APs then send the result to vehicles within the requiring areas. Since the edge server and APs are connected by high-speed wired links whose transmission rate is much higher than the transmission rate between the APs and the vehicles, the transmission delay between the edge server and the APs is ignored.

3.5.2. Processing Delay

Queuing occurs when multiple tasks arrive at an edge server and the edge server has only one service queue with a queuing model of $M/M/1$ [33]. The average waiting time for the tasks in the edge server is

$$t_{w_{\mathrm{com}}}\left(\tau \right)=\frac{{\lambda }_s}{{\mu }_s\left({\mu }_s-{\lambda }_s\right)}$$

(10)

where ${\lambda }_s$ and ${\mu }_s$ represent task arrival rate and the service rate of the edge server, respectively.

The processing time of task ${\tau }_{i,j}$ on the edge server can be written as

$$t_{\mathrm{com}}\left(\tau \right)=\frac{l_{\tau }C}{F}$$

(11)

where the period required by the edge server to complete every bit of task is C, and F is the computational frequency of the edge server.

Thus, the total offloading delay of task $\tau$ is the sum of the transmitting delay and the processing delay, i.e.,

$$t_{\mathrm{total}}\left(\tau \right)=t_{w_{\mathrm{up}}}\left(\tau \right)+t_{\mathrm{up}}\left(\tau \right)+t_{w_{\mathrm{com}}}\left(\tau \right)+t_{\mathrm{com}}\left(\tau \right)+t_{\mathrm{down}}\left(\tau \right)$$

(12)

3.6. Optimization Formulation

When offloading a perception task, two optimization goals should be achieved. The first goal is to minimize the average offloading delay and the second goal is to maximize the shareability factor. The optimization goals are expressed as follows.

$$\begin{array}{cccc}\hfill & \underset{}{\min }\hfill & \hfill & \frac{{\sum }_{\tau \in \mathrm{O}}{t}_{\mathrm{total}}\left(\tau \right)}{\left|\mathrm{O}\right|}\hfill \\ \hfill & \mathrm{s}.\mathrm{t}.\hfill & \hfill & 0\le f_i\le F\hfill & \hfill \left(\mathrm{C}1\right)\\ \hfill & \hfill & & t_{\mathrm{total}}\left(\tau \right)\le \Delta \hfill & \hfill \left(\mathrm{C}2\right)\\ \hfill & \hfill & & {\lambda }_s\le {\mu }_s\hfill & \hfill \left(\mathrm{C}3\right)\end{array}$$

(13)

$$\begin{array}{cccc}\hfill & \underset{}{\max }\hfill & \hfill & \frac{{\sum }_{\tau \in \mathrm{O}}{k}_{\tau }}{\left|\mathrm{O}\right|}\hfill \\ \hfill & \mathrm{s}.\mathrm{t}.\hfill & \hfill & k_{\tau }\in {\mathbb{Z}}_{+}\hfill & \hfill \left(\mathrm{C}4\right)\\ \hfill & \hfill & & C1,C2,C3\hfill & \hfill \end{array}$$

(14)

where C1 represents that the computing capacity allocated to each perception task is lower than the total computing capacity of the edge server; C2 requires that the total latency must not exceed the maximum tolerable latency of each perception task; C3 means that the service capability of the edge server is not less than the arrival rate of perception tasks; and C4 indicates that it is an integer constraint.

4. Methods

4.1. Offloading Based on Greedy Sharing Strategy

In order to decide which tasks are offloaded to the edge servers, an offloading strategy based on greedy sharing, referred to as OG, is designed. The input is the task set T generated by the sensors from the four directions of a vehicle. For a task from a certain direction, the requiring areas are defined as demonstrated in Figure 4 and Figure 5. The edge server assigns a weight $w_{\tau }$ to each task according to the number of vehicles in the requiring areas of the task. The larger number of vehicles in the requiring areas means a larger weight. The edge server determines the offloading order of the tasks based on their weights. The task with a larger $w_{\tau }$ is offloaded with a higher priority to ensure that the task result can be shared with as many vehicles as possible. The process of offloaded task selection is described in Algorithm 1.

Algorithm 1 Offloading Based on Greedy Sharing Strategy (OG).

Input: Task set ${\mathrm{T}}_{\mathrm{all}}$ and computing resource F of edge server. Output: The set of offloaded tasks $\mathrm{O}$. 1: Initialize the set of offloaded tasks $\mathrm{O}=\varnothing$; 2: Assign weights $w_{\tau }$ to each perception task according to Equation (1); 3: Sort task $\tau \in {\mathrm{T}}_{\mathrm{all}}$ in a descending order by $w_{\tau }$; 4: for each $\tau \in {\mathrm{T}}_{\mathrm{all}}$ do 5:       ${\mathrm{c}}_{\tau }=l_{\tau }\ast C$; 6:       if ${\mathrm{c}}_{\tau }<\mathrm{F}\&\&t_{\mathrm{total}}\le \Delta$ then 7:             $\mathrm{O}=\mathrm{O}+\left\{\tau \right\}$; 8:             $\mathrm{F}=\mathrm{F}-c_{\tau }$; 9:       else if $F==0$ then 10:           break; 11:     end if 12: end for 13: return $\mathrm{O}$;

In Line 1, the set of tasks to be offloaded is initialized. In Line 2, the weights of the tasks are determined according to Equation (1). If the result of the task is required by more CAVs, it has a higher weight. In Line 3, the tasks are sorted in a descending order by weight. Hence, a task with higher a weight has a higher priority to be offloaded. From Line 4 to Line 12, the set of offloaded tasks using greedy strategy is obtained. More specifically, for each task (Line 4), the needed resource of task $\tau$ is computed in Line 5. If the edge server can afford the task and the delay of the task is tolerable (Line 6), the task is added into the offloaded set $\mathrm{O}$ (Line 7). After that, the remaining resource of the edge server is updated (Line 8). If no more resource is available, no more computation is needed (Lines 9 and 10). In Line 13, the set of selected tasks to be offloaded is returned.

The time complexity of Algorithm 1 is determined by sorting the tasks (Line 3), which is $O\left(\right|{\mathrm{T}}_{\mathrm{all}}|log|{\mathrm{T}}_{\mathrm{all}}\left|\right)$.

4.2. Offloading Based on Greedy Sharing Strategy with Load Balance

The above offloading algorithm only considers the task weight as the basis of the offloading decision. However, when the task weights of different vehicles are of a great difference, a phenomenon may occur that all tasks of a vehicle are offloaded, while another vehicle has to process all the perception tasks by itself. Such unbalanced offloading may result in significant perception delays for those vehicles that cannot offload their tasks to the edge server.

Algorithm 2 is an improved version of Algorithm 1. In Line 1, the set of offloaded tasks is initialized as an empty set. For each vehicle (Line 2), its perception tasks are assigned weights (Line 3), and these tasks are sorted in an ascending order by weight (Line 4). In Line 5, the task set ${\mathrm{T}}_i$ of the i-th vehicle is split into two subsets. One is denoted as ${\mathrm{T}}_{vi}$, containing the tasks that can be processed by the vehicle itself. The other subset is denoted as ${\mathrm{T}}_{si}$, containing the tasks that have to be offloaded to the edge server. From Line 7 to Line 9, the tasks in ${\mathrm{T}}_{si}$ are added into the offloaded task set $\mathrm{O}$. In Line 10, the tasks in each ${\mathrm{T}}_{vi}$ are put into a candidate set $\Omega$, and are sorted in descending order. From Line 11 to Line 19, each task in $\Omega$ is checked and offloaded if the edge server can bear it.

The computational complexity of Algorithm 2 is also determined by sorting the tasks (Lines 4 and 10), which is $O\left(\right|{\mathrm{T}}_{\mathrm{all}}|log|{\mathrm{T}}_{\mathrm{all}}\left|\right)$.

Algorithm 2 Offloading Based Greedy on Sharing Strategy with Load Balance (OG-LB).

Input: Task ${\mathrm{T}}_{\mathrm{all}}$, computing resource of edge server F, the number of tasks that a vehicle can deal with at most in a perception cycle $\epsilon$ Output: The set of offloaded tasks $\mathrm{O}$. 1: Initialize the set of offloading tasks $\mathrm{O}=\varnothing$; 2: for each ${\mathrm{T}}_i$ do 3:       Assign weight $w_{\tau }$ to each task $\tau \in {\mathrm{T}}_i$; 4:       Sort the tasks in an ascending order by $w_{\tau }$; 5:      Split ${\mathrm{T}}_i$ into two subsets ${\mathrm{T}}_{vi}$ and ${\mathrm{T}}_{si}$, where ${\mathrm{T}}_{vi}$ contains $\epsilon$ tasks with the least weights, and ${\mathrm{T}}_{si}$ contains the remainder $(4-\epsilon )$ tasks; 6: end for 7: for each ${\mathrm{T}}_{si}$ do 8:       $\mathrm{O}=\mathrm{O}\cup {\mathrm{T}}_{si}$; 9: end for 10: Let candidate tasks $\Omega =U{T}_{vi}$, and sort tasks $\tau \in \Omega$ in a descending order by $w_{\tau }$; 11: for each $\tau \in \Omega$ do 12:       ${\mathrm{c}}_{\tau }=l_{\tau }\ast C$; 13:       if ${\mathrm{c}}_{\tau }<\mathrm{F}\&\&t_{\mathrm{total}}\le \Delta$ then 14:             $\mathrm{O}=\mathrm{O}+\left\{\tau \right\}$; 15:             $\mathrm{F}=\mathrm{F}-c_{\tau }$; 16:       else if $F==0$ then 17:             break; 18:       end if 19: end for 20: return $\mathrm{O}$;

5. Results and Discussion

5.1. Experimental Settings

In order to evaluate the performance of the proposed algorithms, simulation experiments are conducted using Python script and SUMO (Simulation of Urban MObility) as the experimental platform. The coverage range of the AP is set to 200 m. The length of the requiring area is set to 100 m, and its width is set to 4 m (equal to the lane width). The edge server computing frequency is 8 GHz. The channel model follows the Rayleigh fading model. The detailed parameters of the experimental settings are listed in

Table 1. Experimental parameters.

Parameter	Value	Denotation
Bandwidth for V2I	40 MHz [12]	$B_{v2i}$
Bandwidth for V2V	10 MHZ [34]	$B_{v2v}$
Transmission power for uplink	23 dBm [28]	$P_r$
Transmission power for downlink	33 dBm [28]	$P_d$
Transmission power for V2V	10 dBm [35]	$P_v$
Noise power	−100 dBm [12]	N
Path loss index	3	$\alpha$
Size of task	1–3 MB	l
Ratio of result size to task size	0.001–0.01	$\beta$
Maximum tolerable latency	100 ms [36]	$\Delta$
Server computing frequency	8 GHz [37]	F
Compute capacity	10 period/bit	C
Simulation scenario	straight; intersection
The coverage range of an AP	200 m
The size of a requiring area	100 m × 4 m
The size of traffic flow	$\{1000,1250,1500,1750,2000\}$ veh/h

.

Each contrast scheme contains two issues, i.e., the offloading strategy and the task result sharing scheme.

For the offloading strategy, the following three strategies are compared:

• Offloading based on greedy algorithm (denoted as OG). As shown in Algorithm 1, the tasks are selected to be offloaded based on the weight $w_{\tau }$.
• Offloading based on greedy algorithm with load balance (denoted as OG-LB). As shown in Algorithm 2, the computation load of each vehicle is taken into account.
• Offloading based on the generation order in the vehicle (denoted as OV). In this scheme, the shareability of the tasks are not considered.

For the task result sharing scheme, the following two schemes are compared:

• The task result is shared by the server (denoted as SS). This is our proposed scheme. In this scheme, the result of the offloaded task is broadcast by the server to all vehicles in the requiring areas through infrastructure-to-vehicle (I2V) communications.
• The task result is shared by the vehicles (denoted as SV). This is a traditional cooperative perception scheme in vehicular edge computing. In this scheme, the edge server sends the result of the offloaded task to the task-generated vehicle, then the vehicle sends the result to other vehicles through vehicle-to-vehicle (V2V) communications.

Hence, a contrast method is denoted as a combination of the the offloading strategy and the task result sharing scheme, such as OG-SS or OV-SV.

To facilitate the representation of the result delivery delay after offloading a task, an experimental variable (denoted as $\beta$) is introduced as the ratio of the task result size to the originally generated task size.

5.2. Results Analysis

5.2.1. Delay

In the simulation experiment, a task is generated for each direction of each vehicle in a perception cycle. In Figure 6, the solid line indicates the SS scheme and the dashed line represents the SV scheme. Obviously, the average delay increases substantially with $\beta$ in the SV scheme, while the average delay hardly varies with $\beta$ by applying the SS scheme, which generally reduce the average delay by 5 ms. Under the same condition, the delay obtained in the OG offloading algorithm is lower than in the OV offloading algorithm.

Since the size of a task is assumed to be fixed, the varying of $\beta$ means a change in the size of the task result. When the task results are delivered using the sharing strategy by the edge server, it is broadcast to all vehicles in the requiring area by the APs. When the size of the task result increases, the delivery delay of the SS scheme only grows at a low speed.

We also compared the numerical relationship between the number of vehicles and the average delay under these offloading strategies, as shown in Figure 7. The ratio of the result size to the task size is set to 0.05, and the number of vehicles varies from 50 to 100. With the increase in vehicles, the average delay also increases. When the number of vehicles reaches 100, the average delay of the OV-SV method exceeds 100 ms, which may endanger driving safety. In contrast, our proposed method (OG-SS) has the lowest delay, which is about 20% lower compared to that of the traditional method (OV-SV). Generally, the average delay of the SS scheme is obviously lower than that of the SV scheme, and the delay obtained from the OG offloading algorithm is about 10 ms less than the OV offloading algorithm.

5.2.2. Shareability Factor

The shareability factor of the perception task represents how many vehicles need the result of the task. Hence, the higher the shareability factor is, the higher the efficiency achieved. For comparison with the traditional method (i.e., OV-SV), the value of the shareability factor in OV-SV is set to be a constant 1. The shareable feature of the task result is investigated in the intersection scenario and the straight road scenario.

The value of the shareability factor is closely related to traffic flow (in veh/h). Traffic flow is set to 1000–2000 veh/h, and the simulation steps are set to 3600 (representing an hour). According to Figure 4 and Figure 5 in Section 3.3, the number of vehicles in the requiring areas is recorded every 200 steps.

The relationship between the shareability factor and the traffic flow in the intersection scenario is shown in Figure 8. Since it is assumed in Algorithm 2 (OG-LB) that a CAV can process $\epsilon$ $(0<\epsilon \le 4)$ tasks in a perception cycle, the algorithms are compared under four different values of $\epsilon$. The four subplots in Figure 8 correspond to the values $\epsilon =1$, $\epsilon =2$, $\epsilon =3$, and $\epsilon =4$, respectively. Three observations can be found in Figure 8. First, as the traffic flow becomes higher, the shareability factors of methods OG-SS, OG-LB-SS, and OV-SS all increase, and OG-LB-SS has the fastest growth rate in most cases. For OG-LB-SS, the shareability factors reach 17, 12, 11.9, and 7.2 when $\epsilon$ is set to 1, 2, 3, and 4, respectively. The reason is that the gain of a sharing task result is larger when the traffic flow is higher. Second, when $\epsilon$ is small, OG-LB-SS has obvious advantages. This is because a smaller $\epsilon$ means a lower processing power of the vehicle, and the effect of task offloading with load balance is more remarkable. OG-LB-SS degenerates into OG-SS when $\epsilon =4$. Third, the shareability factor of OV-SS is smaller than that of OG-SS and OG-LB-SS because it does not guarantee that tasks with more requiring vehicles are offloaded to the edge server in the OV-SS strategy.

Figure 9 depicts the variation of the shareability factor for different traffic flows in a straight road scenario. The four subplots in Figure 9 are also related to the values $\epsilon =1$, $\epsilon =2$, $\epsilon =3$, and $\epsilon =4$, respectively. Similar observations can be obtained with Figure 8. In Figure 9, the method OG-LB-SS also has the fastest growth rate in most cases, and the shareability factors are 10.2, 7.8, 5.3, and 5.0 when $\epsilon$ is set to 1, 2, 3, and 4, respectively. The reasons are similar to the analysis in Figure 8.

When comparing Figure 8 and Figure 9, it can be seen that the shareability factor in the intersection scenario is higher than that in a straight road scenario, since there are more vehicles requiring the result of a certain perception task.

6. Conclusions

In this paper, the shareability of the perception task results was considered for cooperative environmental perception, and a novel vehicular edge computing scheme was designed to make full use of the shareability. In the scheme, the shareability factor of the task result was proposed as a new metric, and the tasks were selected to be offloaded based on this metric. Two versions of the task offloading strategies were designed. One was based on a greedy algorithm, and the other one took load balance into consideration. The edge server then delivered the task results to the CAVs in the requiring areas directly, which reduced the result delivery latency. Simulation results showed that our proposed task offloading scheme outperformed traditional schemes in terms of the task result delivery delay and the shareability factor. Based on the proposed scheme, the efficiency of cooperative perception among CAVs was improved remarkably.

The limitations of this research work include the following: (1) The CAVs in adjacent positions may share similar environmental information, which causes transmission redundancy and (2) only a simulation is conducted to verify the proposed scheme.

In the future, we will further explore a scheme to select a portion of CAVs for cooperative environmental perception based on their real-time dynamic features, such as trajectories, speeds, and positions. Hence, the information redundancy will be reduced, and the efficiency of cooperative environmental perception will be improved greatly. A prototype will also be set up in a real transportation context for verification.