Xiaofei Yu, Chaowei Wang, Lexi Xu, Celimuge Wu, Ziye Wang, Yizhou He, Weidong Wang
©SHUTTERSTOCK.COM/CHESKY
The metaverse employs globally distributed computing and communication infrastructures to construct an immersive digital world. Its continuous synchronization and hyperinteractivity create a dilemma involving tremendous volumes of sensory data and scarce spectrum resources. Connected and automated vehicle (CAV) networks integrate onboard sensing, communication, computation, and storage capabilities to enhance the metaverse. This article introduces an edge intelligence-based mobile crowdsensing (MCS) CAV framework, which studies both perception and transmission dimensions. The metaverse’s cornerstone is high-quality edge sensory data delivery across a geographical distribution. Thus, we construct a CAV crowdsensing-based traffic coverage model. Furthermore, we provide information silos in urban transportation networks as a use case. Simulation results validate the proposed framework’s superiority in improving MCS coverage and perceptual data offloading efficiency. With edge intelligence, such a framework can illuminate the prospect of the convergence between CAV applications and the metaverse.
The science fiction novel Snow Crash first coined the metaverse concept, which has recently regained popularity. Regarded as the “next-generation Internet,” the metaverse allows people to interact in real time via digital avatars in a 3D Internet world [1]. By contrast, in the current mobile Internet era, we utilize a web browser or software via a cursor to navigate the Internet. Indeed, emerging technologies, such as virtual reality/augmented reality (AR)/extended reality (XR), artificial intelligence (AI), and 6G networks, play a crucial role. They enable the construction of a panoramic, immersive, embodied, and interoperable metaverse [2]. Nonetheless, the existing lightweight metaverse does not enable the entire vision of a synchronized virtual world seamlessly connected to the physical world, especially considering the accessibility of worldwide users.
Driven by the paradigm shift from “content inside the screen” to “users inside the content,” the metaverse’s performance relies heavily on collecting and processing data that capture or describe surrounding changes. Such data can enhance renderings and further blur the boundaries between the physical world and digital entities. Specifically, further integration of sensing and communication is critical to guarantee the stable bidirectional data mapping between physical and virtual worlds. As multifaceted Internet of Things objects, CAVs are not only equipped with onboard sensors (radar, cameras, and location receivers) but also with storage, computing, and communication abilities [3]. These sensors and capabilities can help support and enhance fine-grained metaverse applications. Examples include real-time 3D object (e.g., pedestrians, vehicles, and landmarks) detection in AR-based scene sensing for urban driving, which has already been applied in reality. The Internet of Vehicles (IoV) consists of CAVs and roadside infrastructure. By tracking and perceiving environmental information, it can achieve intelligent decision making based on context awareness. Therefore, as shown in Figure 1, the convergence of the metaverse and CAVs enables efficient edge resource management and guarantees the quality of service (QoS) for diverse applications in the metaverse.
Figure 1 The fusion of the metaverse and CAVs.
As a data perception network, MCS involves utilizing sensors on mobile units to perform large-scale sensing tasks. MCS has three main characteristics: 1) a random distribution of sensing nodes and wide sensing area; 2) a high robustness of the sensing network; and 3) a human-centric nature, with humans acting as sensing objects, device managers, data sources, and service recipients. Such characteristics coincide with the metaverse’s features, which include distributed computing and communication infrastructures and humans as the primary players. Thus, such a metaverse context allows MCS to monitor and gather bidirectional mapping information [4]. This information consists of the surrounding information in the physical world. It also includes data about user interactions and experiences in the virtual environment. The above perception data can optimize metaverse network settings and resource allocation in a multidimensional and fine-grained manner. In addition, the CAV network’s scalability, distributed mobile nature, and reliable wireless transmission capability make it the preferred sensing participant for MCS [3]. Figure 2 is a schematic diagram of the MCS system based on CAVs in the metaverse. Multiple CAVs serve as sensing nodes in the MCS network. They receive metaverse application sensing tasks posted on the platform by different users and selectively participate in collaborative task completion with budgetary incentives. In turn, the platform decides which intersections roadside units (RSUs) will be deployed at based on the locations and expected trajectories submitted by the participating CAVs. The deployed RSUs receive edge sensory data opportunistically offloaded by the CAVs.
Figure 2 The MCS system based on CAVs in the metaverse. RSU: roadside unit; BS: base station.
To satisfy metaverse applications’ differentiated user-oriented requirements, the CAV network in the metaverse necessitates enhanced per-vehicle intelligence and cooperation among the participants (vehicle to vehicle and vehicle to RSU). Edge intelligence, combining edge computing and AI, enhances the CAV network’s responsiveness and optimizes resources for demanding tasks in the metaverse. Particularly, the edge for AI brings perception, communication, and AI inference near the data source to support metaverse peer-to-peer interactions. AI for the edge can improve the network-level autonomy of MCS CAV systems with AI algorithms [5]. Fortunately, the metaverse inherits the mobile Internet’s underlying architecture and protocol suite, which allows all sensing and transmission technologies to be applied [1]. Moreover, in reality, the data derived from various vehicle brands and sensors exhibit heterogeneity. Ensuring edge inference consistency and sensory data availability across the geographic distribution is vital. Therefore, an urgent reorchestration of the decentralized perception service delivery CAV framework is necessary for efficient information exchange and optimization of the system’s generalization capability.
Preliminary works have demonstrated that MCS is an efficient, reliable, and scalable approach to acquiring data for diverse intelligent applications. It takes mobile devices as sensing terminals to perceive, gather, and transmit large-scale data through collaborative wireless communication. It outperforms traditional static wireless sensor networks, especially regarding deployment and maintenance costs, sensing range and granularity, and resilience [6], [7], [8]. However, these works focus more on the single aspect of MCS issues and hardly balance the metaverse ecosystem requirements.
This article aims to motivate the metaverse implementation by proposing an edge intelligence-based MCS CAV framework. Specifically, the MCS CAV framework employes AI algorithms to create customized recruitment strategies for sensing participants in various metaverse applications and dynamic network topology shifts. To achieve a real-time digital duplication of the physical world, our initial work [9] constructs a CAV crowdsensing-based traffic coverage model, which grids urban districts and introduces the concepts of the region of interest (ROI) and spatiotemporal coverage (STC). Therefore, the tasks performed can obtain heterogeneous perceptual data with spatiotemporal distribution properties. However, this study highlights the MCS CAV framework’s perception perspective and neglects the exploration of network transmission and data offloading. To facilitate collaborative services provided by participating CAVs and minimize redundant overhead from isolated vehicles’ repetitive trajectories, the MCS CAV framework designs a deep reinforcement learning (DRL)-based strategy for selecting participants. This approach enhances perceptual data quality and alleviates the oversized CAV selection space problem. On this basis, the MCS CAV framework executes the RSUs’ deployment layout for opportunity-based data offloading to mitigate traffic congestion and delay problems in the transmission network. Through the fusion of the metaverse and MCS, combined with CAVs equipped with various sensors, users can manifest in different virtual immersive landscapes. Hence, jointly optimizing the framework configuration from both the perception and transmission perspectives can improve the edge sensory data quality while coping with the latency caused by spectrum occupation and ensuring the QoS of applications.
The remainder of the article is structured as follows. We first introduce the MCS CAV framework hierarchy and its corresponding architecture functions. Then, we present a CAV crowdsensing-based traffic coverage model. We also go through a case study of solving the information silos problem in urban transportation networks via the proposed MCS CAV framework and model. Extensive simulations are performed to investigate the approach’s superiority. Finally, we conclude the article, followed by a discussion of future directions.
This section presents a framework overview of the edge intelligence-based MCS CAV system. Furthermore, its hierarchical architecture and corresponding functions are also introduced.
A typical MCS system consists of multiple task requesters, many mobile participants, and a cloud MCS platform. The requester submits a perception task to the platform based on the service requirements and its conditions. The platform then disseminates and recruits participants for this task, with a reward for their contributions. Eventually, the task participants with devices perform this task and upload data to provide services. Therefore, with reference to the classic structure and workflow of MCS, we propose a hierarchical framework of the edge intelligence-based MCS CAV system in the metaverse, as described in Figure 3.
Figure 3 The hierarchical architecture, modules, functions, and scenarios for the edge intelligence-based MCS CAV framework in the metaverse.
There are three main modules in this framework: the service requester (i.e., metaverse-supported immersive services), the service supporter (i.e., digital twin-supported metaverse functions), and the intelligent CAV workers (i.e., MCS CAV-supported physical space). This architecture works from the bottom up, sequentially from the perception layer for data collection to the service layer for data enablement. These layers iteratively perform mapping and computation operations to achieve modules’ coupling. In particular, we can optimize the perception and transmission layers’ physical resources to ensure the data quality required by the upper-layer services. The advantages of this framework are as follows:
The layered architecture of the edge intelligence-based MCS CAV system in the metaverse has six layers in a bottom-up manner. Each layer and its corresponding functions are described as follows.
The perception layer consists of two functions: perception information acquisition and preprocessing. Before executing a task, the CAVs, as edge intelligent workers, should identify the perception location and rate depending on the sensing task content and their trajectories. They perform a simple preprocessing operation to ensure data freshness after the sensory information has been successfully gathered.
The link layer is responsible for building the channels for user access and network communication. It also provides reliable data transmission links for the upper-layer services.
As the highest layer of communication-oriented functions, the transmission layer provides reliable end-to-end connectivity. The system should design spectrum control and access policies based on constrained wireless resources to ensure sensed data quality during transmission.
Driven by the collected sensed data, the application layer can provide appropriate intelligent services to task requesters. Meanwhile, participating CAVs can be rewarded based on their contributions.
The functionality layer, supported by digital twin technology, provides executability assessment and state awareness for the practical business. It can construct models based on the lower-layer data according to the requirements and provide the corresponding functions for the upper-layer business.
The service layer is a straightforward user-oriented layer with the users’ quality of experience (QoE) as the primary indicator. Its purpose is mainly to provide users with an immersive environment for good human–computer interactions. Users can easily access the information the function layer provides for decision making via voice or physical gestures.
Consequently, the MCS CAV framework adeptly oversees the entire construction process of the metaverse, spanning from its foundational layers to the upper layers, thereby ensuring the holistic quality of the perception data. Simultaneously, fine-tuning physical resources within the perception and transmission layers assures the real-time efficacy of the perception data and the reliability of the edge intelligence inference.
This section constructs and analyzes a CAV crowdsensing-based traffic coverage model. We also adopt the concepts of the ROI and STC to characterize the model better.
In the MCS CAV framework, participants and data are two critical considerations. From the participant perspective, CAVs work as edge intelligent executors. MCS can leverage the CAVs’ wide presence, flexible mobility, and opportunistic connectivity to perform perception tasks. From the data perspective, MCS can mine and exploit the spatiotemporal correlation of CAV-perceived data via edge intelligence algorithms to improve data utilization efficiency. Thus, according to the proposed edge intelligence-based MCS CAV framework, this article abstracts the MCS CAV-supported physical space and constructs a CAV crowdsensing traffic coverage model, as in Figure 4.
Figure 4 The CAV crowdsensing-based traffic coverage model.
This article defines the ROI as a lattice geographic area containing several road segments and some urban areas. The urban area in Figure 4 is divided into four grid cells, ROI 1, ROI 2, ROI 3, and ROI 4, respectively. It is worth mentioning that we also provide a demonstration of the ROI and CAVs’ continuity trajectories combination in Figure 4, where the dots on the trajectory indicate CAV locations. Regardless of the vehicle location at the ROI center or the ROI edge, each CAV is supposed to perceive all such ROI information during the perception period. We assume that the set of CAVs is V. Therefore, as the basic sensing unit to measure the perception coverage, no matter where a CAV drives, its position will always be in one of the ROIs, and it can gather the information in that ROI. The set of trajectories is also the set R of ROIs passed by.
For the task requester, the target area for data collection always consists of several ROIs, which can be spatially continuous or dispersed. In a nutshell, the urban environment information collection by the MCS CAV system is equivalent to the information collection of different ROIs. Therefore, the number of ROIs can measure the MCS network’s coverage degree for the city. In addition, combining the data’s temporal characteristics, continuous time is discretized into several perception periods and forms the set T. The MCS platform collects the information sensed by the CAVs once every time interval $\Delta{t}$. Therefore, according to the sets V, R, and T, we can utilize a matrix to represent the set of vehicle trajectories: \begin{align*}{L}\left({{V},{T}}\right) &= \left[{\begin{array}{cccc}{{l}\left({{v}_{1},{t}_{1}}\right)}&{{l}\left({{v}_{1},{t}_{2}}\right)}&{\cdots}&{{l}\left({{v}_{1},{t}_{n}}\right)}\\{{l}\left({{v}_{2},{t}_{1}}\right)}&{{l}\left({{v}_{2},{t}_{2}}\right)}&{\cdots}&{{l}\left({{v}_{2},{t}_{n}}\right)}\\{\vdots}&{\vdots}&{\vdots}&{\vdots}\\{{l}\left({{v}_{m},{t}_{1}}\right)}&{{l}\left({{v}_{m},{t}_{2}}\right)}&{\cdots}&{{l}\left({{v}_{m},{t}_{n}}\right)}\end{array}}\right] \\ {l}\left({{v}_{i},{t}_{j}}\right)&\in{R},{v}_{i}\,{\in}\,{V},{t}_{j}\,{\in}\,{T}{.} \tag{1} \end{align*}
According to the vehicle trajectory matrix ${L}\left({{V},{T}}\right)$, we assume that the CAVs number is m and that the sensing periods number is n. Ideally, the total spatial coverage during such a duration is ${m}\,{\times}\,{n}$ ROIs. However, if there are multiple CAVs in the same ROI simultaneously, the data collected by these CAVs will overlap. Although having a certain degree of redundancy in sensory data can ensure reliability, the system must also reduce the overlap of trajectories to improve the efficiency of physical resource utilization. Therefore, we introduce STC to characterize the degree of trajectory overlap. It is defined as the total number of different ROIs covered during the whole perception duration and contains both temporal and spatial coverage dimensions. We describe its calculation process as follows:
In this section, we evaluate the performance of our proposal with an example of the information silos problem in urban transportation under the proposed MCS CAV framework in the metaverse. We respectively optimize the perception and transmission layers’ resources in the MCS CAV-supported physical space to support diverse immersive metaverse applications. In particular, a deep Q-network (DQN)-based sensing participants’ selection strategy is applied in the perception layer, and an RSU-based opportunistic CAV data offloading optimization algorithm is adopted in the transmission layer.
The development of XR, the IoV, and wireless communication technologies has made the metaverse’s realization conceivable, which has likewise caused the sensing participants’ scale and the perception data’s volume to be much larger. However, the data types’ diversity and multiple sources, the data capacity’s massiveness, and the compartmentalization between different components all put growing pressure on the data processing and fusion at the server [10]. Especially in the urban transportation information management scenario, it is difficult and inefficient to share information among CAVs. All these factor contribute to worsening the information silos problem in urban transportation. Therefore, we study the MCS participant recruitment problem and the MCS opportunistic data offloading RSU deployment problem in an urban transportation network scenario.
In real-life scenarios, certain predictability exists in the CAVs’ daily routines. Hence, the traffic flow exhibits periodic and spatiotemporal traits, underscoring the vital role of tracking CAV trajectories. This tracking is critical for strategically choosing CAVs to engage in sensing tasks. However, CAVs will consume certain costs, such as power, oil usage, and so on, in performing perception tasks. They do not participate in the tasks published by the platform unconditionally. Therefore, the MCS platform should offer some incentives to encourage CAVs to engage in the perception tasks issued actively. At the same time, only CAVs that admit this incentive will be candidates for such a task. However, no matter what incentive is designed, the task requester will have a total reward cap, i.e., the perception budget. The total reward of the selected CAV workers cannot exceed the perception budget. Therefore, this strategy aims to select the appropriate CAV set to maximize the STC without exceeding the perception budget. Considering this problem’s NP-complete characteristics and the vehicle driving routines, we can model the sensing CAV participants’ selection problem as a Markov decision process (MDP) and obtain the target set using a DRL algorithm.
Specifically, based on the proposed CAV crowdsensing-based traffic coverage model, we model the interactions between the MCS server and the environment as an MDP. In particular, the environment, in this case, refers to the urban transportation network, and the state is a property of the environment. The state in this context is designed as the CAVs’ trajectory information, i.e., (1). The MCS server acts as an agent by observing the current vehicle driving trajectory and applying an action, i.e., selecting a set of CAVs. Then the environment transfers into the next state and returns the reward to the agent. The agent keeps adjusting the action for the following iteration based on the reward. As a classic example of DRL, the DQN algorithm can integrate Q-learning with deep learning to avoid an excessive state and action space in the proposed scenario [11]. Meanwhile, scheduling under a unified MCS platform alleviates the problem of segmentation among different components and data heterogeneity in urban transportation information silos. Hence, we illustrate the proposed DQN-based sensing CAV participants’ selection algorithm in the following three aspects:
A specific execution flowchart of the algorithm appears in Figure 5. The environmental parameters and system state are fed into the network ${Q}\left({{s}_{t},{a}_{t}{;}{\omega}}\right)$, which is then trained to approximate ${Q}\left({{s}_{t},{a}_{t}}\right)$. The output experience data are stored in a replay memory with a capacity of ${N}_{B}$. The data in the replay pool are a quadruplet containing the current state ${s}\left({t}\right)$, the action ${a}\left({t}\right)$, the immediate reward ${r}\left({t}\right)$ obtained by the action, and the next state ${s}\left({{t} + {1}}\right)$ to be transferred to. The algorithm randomly samples a minibatch of experience data from the replay memory at each iteration and trains the Q-network to update the parameters ${\omega}$. The algorithm uses an $\varepsilon{-}{\text{greedy}}$ policy for action selection; $\varepsilon$ indicates that the algorithm has a probability of $\varepsilon$ to randomly select an action at each state transition. It also has a probability of ${1}{-}\varepsilon$ to select the action that results in the best Q value in the current state. We set the $\varepsilon$ larger at the beginning of the training process to explore the action space more thoroughly. During the training process, the exploration rate decreases linearly from the initial “${\varepsilon}_{s}$” to the end “${\varepsilon}_{e}$” in a specific increment.
Figure 5 The DQN-based sensing participants’ selection algorithm.
Most services the metaverse provides are relevant to the context of the surrounding environment. These context-aware services can leverage the storage capacity of roadside infrastructure to improve users’ QoE and efficiently allocate network resources. Meanwhile, an increase in the number of sensing CAV participants and the quality requirements of the sensing data both generate huge data volumes. Moreover, the frequent communication of perception data over cellular networks will cause congestion and occupy authorized spectrum resources. Opportunistic data offloading uses roadside infrastructure to provide CAVs with high-speed, stable, and affordable Wi-Fi networks to offload perception data and further cache such data, which makes it a reliable content delivery solution under the metaverse MCS CAV framework.
In the proposed CAV crowdsensing-based traffic coverage model, RSUs deployed at traffic intersections are the ideal infrastructure for data offloading. They act similarly to wireless access points and help CAVs offload data requested from metaverse intelligent applications. We employ wireless gigabit technology to provide point-to-point broadband links with speeds in excess of 4 Gb/s for multiple clients simultaneously. The CAVs follow an opportunistic transmission model called the “store-and-forward” model. Specifically, when a vehicle opportunistically stays at a traffic light or in a parking lot and enters the RSU’s service coverage, point-to-point data offloading is performed through the RSU. The RSU stores and transmits these data centrally back to the MCS server. When the vehicle leaves the RSU service coverage, its data transmission is switched back to the cellular network until the next contact with the RSU.
Within the allowable execution time of perception tasks released by the MCS platform, this strategy adjusts the RSU deployment location or enablement status within the requester’s budget for dynamic CAV network topologies and complex wireless environments. The optimization goals are to improve RSU utilization and reduce system response time. The problem is NP-hard, and we apply an improved greedy algorithm to find the near-optimal solution. Considering the latency tolerance time constraints for different perception tasks, each CAV does not have to send perception data back to the server immediately but can opportunistically offload to the RSU during the time limit. The RSU can cache and uniformly preprocess the offloaded data for transmission. Hence, this strategy alleviates spectrum resource scarcity and insufficient resource utilization in urban transportation information silos.
Specifically, we abstract the urban transportation network as an undirected graph G. The nodes in the graph represent the intersections in the urban roads, and the edges in the graph represent the roads connecting each intersection. Based on this, we can utilize matrices to represent the geographic coordinates of m CAVs at n sensing periods and the deployment coordinates of RSUs selected by the policy. Further, we define the CAVs’ network access state matrix, an ${m}\,{\times}\,{n}$ 0–1 matrix computed from the vehicle and RSU deployment coordinates. The element s in this matrix is a Boolean value, where 1 indicates that the vehicle is connected to the Wi-Fi network provided by the RSU, and 0 is the opposite. Also, we mark C as the service time, which is equal to the sum of the CAV access network status successively to 1. Similarly, the waiting time D is the sum of the CAV access network status successively to 0. A pair of C and D constitutes a CAV’s data collection and offloading phase. The whole perception task execution can be divided into K phases. A schematic diagram of the CAV network access state timeline is in Figure 6. The lost data amount in Figure 6 is the amount of data not uploaded when the CAV leaves the RSU coverage. And these data are discarded when their collection time is beyond the time limit of the perception task.
Figure 6 The network access timeline of CAV i.
This strategy assumes that the RSU utilization rate ${U}_{R}$ and the normalized average service response time ${T}_{w}$ are essential parameters for evaluating offloaded data quality. Here, ${U}_{R}$ is the total amount of data offloaded to the RSU divided by the amount of all the perception data, representing the percentage of the data offloaded through the RSU. In addition, ${T}_{w}\in$(0,1], and ${T}_{w}\propto{\bar{D}}_{i}$. Therefore, this strategy’s optimization objective is the quotient of ${U}_{R}$ and ${T}_{w}.$ Capital costs, maintenance costs, and congestion levels at intersections influence RSU deployment costs. We define the RSU deployment cost set as P, corresponding to the intersection set. In summary, we abstract the RSU deployment problem as selecting a set of suitable RSU deployment locations from the set of intersection coordinates to maximize the optimization objective under the constraint of a deployment budget.
We use an improved greedy algorithm to find this problem’s approximate optimal solution. First, we define the marginal effect, which is the amount of reduction in the objective function by deploying a new RSU compared to the overhead of the current RSU. The improved greedy strategy is to preselect a set of the most efficient intersections for RSU placement and then perform the rest of the greedy iterations. This strategy considers the marginal effect and always selects the RSU with the most significant marginal effect in the current iteration. When the deployment budget is exhausted or the marginal effect tends to zero, the algorithm automatically stops and proposes the current optimal strategy. The improved strategy dramatically reduces the algorithm’s complexity, which is ${O}({mn})$.
This section evaluates the STC and data offloading efficiency by comparing them with listed benchmarks under different parameter settings to verify the feasibility of the adopted scheme.
The metaverse is an immersive digital representation of the physical world, which can be either a traditional environment’s digital mirror or an utterly fictional space [12]. Our proposed edge intelligence-based MCS CAV framework in the metaverse preserves digital copies of the actual CAVs that can perceive and collect information about the surroundings. Therefore, to better validate the algorithm performance, we construct the simulation environment with the help of realistic recorded trajectory datasets, which are the UCI GPS Trajectory database and a Shanghai, China, cab trajectory dataset. Meanwhile, in the metaverse, we use a virtual time system to accelerate the sample collection and transmission of CAVs, avoiding the actual timescale limitation in the physical world.
In the simulation for the DQN-based sensing CAV participants’ selection strategy, we extract the trajectories of 38,902 vehicles from the UCI GPS Trajectory database, collected in the time range of 7:24 a.m. on 13 September 2014 to 1:01 p.m. on 19 January 2016 [13]. This strategy employs a deep neural network with three hidden layers to approximate the nonlinear function. Each hidden layer owns 256, 256, and 512 neurons, respectively. The activation function in the first two hidden layers is a rectified linear unit. In the output layer, the activation function is the tanh function. Also, the replay memory capacity ${N}_{B}$ is 5,000. The minibatch size is 32. The positive and negative reward value ${\lambda}^{+}$ and ${\lambda}^-$ are $ + {1}$ and ${-}{1}{;}$ ${\varepsilon}_{s}$ and ${\varepsilon}_{e}$ are 0.99 and 0.001. For all the CAV candidates, the sensing cost is uniformly distributed in [1] and [5].
We introduce three baseline algorithms for comparison, as follows. The random CAV selection algorithm chooses the sensing CAVs randomly within the constraint of the perceptual budget. The greedy CAV selection algorithm picks the sensing CAVs by maximizing the total reward within the limit of the perceptual budget. The enhanced global congestion awareness (EGCA) algorithm accomplishes the perception task by sequentially selecting the optimal vehicle based on the marginal gain in the STC from the set of candidate vehicles through multiple iterations [9]. Figure 7 compares the STC performance under different selection strategies. The STC increases with the number of CAV candidates because more CAVs bring more trajectory choices, and the MCS server can recruit more suitable CAVs to optimize the final selections. It is worth mentioning that when the perceptual budget consumption saturates, there is no space left to recruit more CAVs. Therefore, at the end of the curves, the rising gradient of the STC decreases and is essentially flat even though the number of CAV candidates increases. We can also observe that the proposed algorithm outperforms other algorithms regarding the STC performance metric because the greedy algorithm focuses only on the total perceived reward satisfaction rather than the STC performance. The random algorithm cannot obtain the optimal solution since the CAV participants are stochastically selected. The EGCA algorithm’s effectiveness is limited by selecting the CAV initial set since it must first randomly select three or more CAVs as the beginning state, causing instability in its performance.
Figure 7 The STC versus the number of CAVs (sensing periods = 10, perceptual budget = 15).
Figure 8 illustrates the STC performance of different algorithms for different sensing periods. The STC increases significantly with the sensing periods. As an essential dimension, the sensing time affects the number of ROIs that each CAV trajectory can cover while performing the perception task. For example, in this experiment, 10 CAVs are involved in the perception task, and it is possible to cover up to 10 more different ROIs for each additional sensing period, resulting in the STC increasing more. The proposed algorithm can obtain the optimal solution and achieve the highest STC compared to the other algorithms.
Figure 8 The STC versus the number of sensing periods (CAVs = 10, perceptual budget = 15).
In the improved greedy RSU deployment strategy simulation, we use a real Shanghai traffic dataset consisting of more than 4,000 cab trajectories [14]. We preprocess the traffic dataset to eliminate inappropriate trajectories. The sampling periods of different data files are scaled to ensure the simulation’s reliability. The encounters between CAVs and RSUs are considered Poisson processes. The number of RSU occurrences in vehicle trajectories obeys a Poisson distribution with ratio “${\beta}$” [15]. The time that a CAV stays within an RSU’s communication coverage is considered an exponential distribution with ratio “${\mu}$.” The deployment cost P follows a Gaussian distribution with a mean of 1. We also employ a fixed signal strength (fixed Really Simple Syndication) channel model for point-to-point transmission, and the data perception rate ${d}_{0}$ and data offloading rate “${\sigma}$” are set to 500 MB/s and 4 GB/s.
We present three baseline algorithms for comparison, as follows. The uniform deployment algorithm places RSUs at intersection coordinates with a fixed interval distance. The classic greedy algorithm always chooses the intersection coordinates with the lowest deployment cost to place RSUs. The single-node capacitated facility location (SNCFL) deployment algorithm adds the RSUs following a decreasing order based on the difference between the total benefits and the deployment overheads [14]. It is worth mentioning that this simulation uses a scaled virtual time clock instead of an actual clock. Figure 9 presents the RSU utilization for different policies with various latency tolerances. The longer the latency tolerance, the higher the RSU utilization. The proposed algorithm’s RSU utilization can approach 100%. Even with a low latency tolerance for real-time applications in the metaverse, about 80% of the data can have a chance to be offloaded through the RSUs. The SNCFL algorithm performs slightly worse in our deployment scenario since it calculates from a benefits perspective. Figure 10 gives the normalized average service response time ${T}_{w}$ versus different deployment budgets; ${T}_{w}$ decreases as the budget increases, and the proposed algorithm always maintains the lowest response time because the rising budget can deploy more RSUs in the urban traffic network to expand the RSUs’ coverage. In turn, the adequate communication time between CAVs and RSUs increases. The proposed algorithm reduces the service response time and enables the RSUs to provide an extended data offloading duration to the CAVs. As a result, the overall data offloading efficiency increases as the deployment budget grows. Also, due to the decrease in ${T}_{w}$, it is almost impossible for data to wait beyond the service’s latency tolerance time, increasing ${U}_{R}$.
Figure 9 The RSU utilization ratio versus the latency tolerance.
Figure 10 The normalized average service response time versus the total deployment budget.
This article proposed an edge intelligence-empowered MCS framework based on CAVs to facilitate the metaverse’s realization. The distributed CAV network’s multidimensional fine-grained perception data can fundamentally support the metaverse to innovate human-centric communication and the Internet. The proposed framework, consisting of three components and six layers, can flexibly adapt to different intelligent applications in the metaverse. We also constructed a CAV crowdsensing-based traffic coverage model that effectively quantifies the state of intelligence, communication, and computation within the physical space. This model evaluates the quality of perceptual data from a spatiotemporal viewpoint. Further, we highlighted the superiority of the proposed framework by implementing a case study of information silos in urban transportation networks. Experiments were conducted to reconcile the perception and transmission layer resource allocation, which uses the DQN algorithm and the application’s delay characteristics in edge intelligence. Simulation results showed that 1) the DQN-based algorithm can obtain the best STC performance compared to the traditional algorithm and provide higher data quality at the perception layer and 2) collaborating with the intelligent application’s latency requirements, the improved greedy algorithm can improve the RSU utilization and provide higher data offloading efficiency at the transmission layer. Thus, the proposed framework can optimize resource allocation with edge intelligence, providing higher data quality and alleviating the spectrum scarcity challenge in the metaverse.
This article is a preliminary exploration of CAV-assisted MCS in the metaverse. With specialized mapping, the proposed framework is expected to apply to metaverse services with increasingly differentiated requirements. Considerable further research is desired in the following areas:
This work is jointly funded by the Open Research Fund of the Shaanxi Province Key Laboratory of Information Communication Network and Security, under Grant ICNS202003, and the Beijing University of Posts and Telecommunications Excellent Ph.D. Students Foundation, under Grant CX2022210. Chaowei Wang and Lexi Xu are the corresponding authors.
Xiaofei Yu (yuxiaofei@bupt.edu.cn) is pursuing the Ph.D. degree in electronics science and technology at Beijing University of Posts and Telecommunications, Beijing 100876, China. She received the B.Sc. degree in electronics science and technology from Beijing University of Posts and Telecommunications in 2018, and she has authored five journal papers. Her research interests include heterogeneous distributed computing, mobile edge computing, and resource management in communication systems.
Chaowei Wang (wangchaowei@bupt.edu.cn) is an associate professor at the School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China. He has authored over 60 journal/conference papers and two books and is the associate editor of Frontiers in Computer Science. His research interests include next-generation mobile communication, wireless sensor, and Internet of Things technology. He is a Member of IEEE and a senior member of the Chinese Institute of Engineers.
Lexi Xu (davidlexi@hotmail.com) is a senior engineer at the Research Institute, China United Network Communications, Beijing 100048, China. He has applied for more than 40 patents, published two books, edited four international conferences proceedings, and held numerous conference roles. His research interests include big data, self-organizing networks, satellite systems, and radio resource management in communication systems. He is a Senior Member of IEEE.
Celimuge Wu (celimuge@uec.ac.jp) is a professor at and the director of the Meta-Networking Research Center, University of Electro-Communications, Tokyo 182-8585, Japan. He is an associate editor of IEEE Transactions on Network Science and Engineering, IEEE Transactions on Green Communications and Networking, and IEEE Open Journal of the Computer Society. His research interests include vehicular networks, edge computing, and artificial intelligence for wireless networking and computing. He is a Senior Member of IEEE.
Ziye Wang (wzy11140@bupt.edu.cn) is pursuing the M.S. degree in electronics science and technology from Beijing University of Posts and Telecommunications, Beijing 100876, China. She received the B.Sc. degree in optoelectronic information science and engineering from Beijing University of Posts and Telecommunications in 2021. Her research interests include mobile edge computing and mobile crowdsensing.
Yizhou He (heyizhou@caict.ac.cn) is a senior engineer at the China Academy of Information and Communications Technology, Beijing 100876, China. He received the Ph.D. degree from Beijing University of Posts and Telecommunications in 2015. His research interests include wireless and mobile satellite communications.
Weidong Wang (wangweidong@bupt.edu.cn) is a professor at the School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China. He is a reviewer for the National Natural Science foundation of China and has authored over 270 journal/conference papers. His research interests include satellite mobile communication, next-generation mobile communication technology, and Internet of Things technology. He is a Member of IEEE and a fellow of the China Institute of Communications.
[1] M. Ball, The Metaverse: And How It Will Revolutionize Everything. New York, NY, USA: Liveright Publishing, 2022.
[2] W. Y. B. Lim et al., “Realizing the metaverse with edge intelligence: A match made in heaven,” IEEE Wireless Commun., early access, Jul. 4, 2022, doi: 10.1109/MWC.018.2100716.
[3] Y. Liu and H. Gao, “Stability, scalability, speedability, and string stability of connected vehicle systems,” IEEE Trans. Syst., Man, Cybern. Syst., vol. 52, no. 5, pp. 2819–2832, May 2022, doi: 10.1109/TSMC.2021.3054794.
[4] W. Wang, Y. Yang, Z. Xiong, and D. Niyato, “Footstone of metaverse: A timely and secure crowdsensing,” IEEE Netw., early access, Mar. 2023, doi: 10.1109/MNET.134.2200598.
[5] C. De Alwis, Q. V. Pham, and M. Liyanage, 6G Frontiers: Towards Future Wireless Systems. Hoboken, NJ, USA: Wiley, 2022.
[6] Z. Dai et al., “AoI-minimal UAV Crowdsensing by model-based graph convolutional reinforcement learning,” in Proc. IEEE Conf. Comput. Commun. (INFOCOM), London, U.K., 2022, pp. 1029–1038, doi: 10.1109/INFOCOM48880.2022.9796732.
[7] Y. Huang et al., “OPAT: Optimized allocation of time-dependent tasks for mobile crowdsensing,” IEEE Trans. Ind. Informat., vol. 18, no. 4, pp. 2476–2485, Apr. 2022, doi: 10.1109/TII.2021.3094527.
[8] H. Wang et al., “Architectural design alternatives based on cloud/edge/fog computing for connected vehicles,” IEEE Commun. Surveys Tuts., vol. 22, no. 4, pp. 2349–2377, Fourth Quart. 2020, doi: 10.1109/COMST.2020.3020854.
[9] C. Wang, X. Gaimu, C. Li, H. Zou, and W. Wang, “Smart mobile crowdsensing with urban vehicles: A deep reinforcement learning perspective,” IEEE Access, vol. 7, pp. 37,334–37,341, 2019, doi: 10.1109/ACCESS.2019.2905263.
[10] W. Xu, Y. Zhang, F. Wang, Z. Qin, C. Liu, and P. Zhang, “Semantic communication for the Internet of Vehicles: A multiuser cooperative approach,” IEEE Veh. Technol. Mag., vol. 18, no. 1, pp. 100–109, Mar. 2023, doi: 10.1109/MVT.2022.3227723.
[11] Z. Zhou, H. Liao, B. Gu, K. M. S. Huq, S. Mumtaz, and J. Rodriguez, “Robust mobile crowd sensing: When deep learning meets edge computing,” IEEE Netw., vol. 32, no. 4, pp. 54–60, Jul./Aug. 2018, doi: 10.1109/MNET.2018.1700442.
[12] Y. Ren, R. Xie, F. R. Yu, T. Huang, and Y. Liu, “Quantum collective learning and many-to-many matching game in the metaverse for connected and autonomous vehicles,” IEEE Trans. Veh. Technol., vol. 71, no. 11, pp. 12,128–12,139, Nov. 2022, doi: 10.1109/TVT.2022.3190271.
[13] M. O. Cruz, H. Macedo, and A. Guimarães, “Grouping similar trajectories for carpooling purposes,” in Proc. Brazilian Conf. Intell. Syst. (BRACIS), Natal, Brazil, 2015, pp. 234–239, doi: 10.1109/BRACIS.2015.36.
[14] Y. Ni, J. He, L. Cai, J. Pan, and Y. Bo, “Joint roadside unit deployment and service task assignment for Internet of Vehicles (IoV),” IEEE Internet Things J., vol. 6, no. 2, pp. 3271–3283, Apr. 2019, doi: 10.1109/JIOT.2018.2882436.
[15] G. Raja, A. Ganapathisubramaniyan, S. Anbalagan, S. B. M. Baskaran, K. Raja, and A. K. Bashir, “Intelligent reward-based data offloading in next-generation vehicular networks,” IEEE Internet Things J., vol. 7, no. 5, pp. 3747–3758, May 2020, doi: 10.1109/JIOT.2020.2974631.
Digital Object Identifier 10.1109/MVT.2023.3320865