Digital Twin Enabled Data-Driven Approach for Traffic Efficiency and Software-Defined Vehicular Network Optimization (2024)

III-A Digital Twin of Roundabout and Connectivity

Fig. 2 illustrates the operation of the proposed DT for an abstracted data driven solution. As input of DT, real-time videos of the target PT are used. These videos are captured using cameras connected to the platform. To detect vehicles in the scenario and assess their states in real-time, we employ a video-based vehicle identification and data extraction method facilitated by a proprietary computer vision software.When a vehicle enters into the video, the software tracks the vehicle and determines the vehicle’s category, location i.e. the lane or approach, speed, and timestamp. The DT system connected with this software using an application programming interface (API), uses this information to create a virtual data object (VO) that simulates the vehicle. As such software cannot classify human driven vehicles (HV) and AV yet, we propose using V2I connectivity to identify CAVs and synchronize their data with the DT-generated VOs based on CAV location. The VO is then inserted into traffic and vehicular networking simulation and updated in real-time until the vehicle leaves the video. This method simulates all the vehicles’ movements and network data flow in DT. Subsequently, critical data such as travel time, waiting time, average speed, queue length, state of connectivity and networking devices, network statistics etc. are extracted from the simulation results and stored in a data repository. From the data-driven solution perspective, this information is used to compare algorithms, control and optimization strategies in real-time, conduct “what-if” analyses, and make predictions in both transportation and vehicular networking domains. To implement each strategy or algorithm, a new DT instance is created, allowing key parameters of vehicles, road infrastructure, CAV models, and vehicular networking and connectivity to be modified. All the instances of the DT run simultaneously and as with the previous process, data is extracted and stored in the repository. Once the data-driven solution performs comparative analysis and selects the optimal strategy or algorithm, it can either transform this into actions to be executed or inform the stakeholders through output message.

III-B Data Driven Decision Support Framework

The decision-support framework of the platform utilizes the data gathered from the PT to make decisions for the CAVs. This framework considers the four approaches of the PT, where the inbound lanes are defined by a set $H$ where,

In equation 1, $N_{in},S_{in},E_{in},\text{and }W_{in}$ represent the inbound lanes of North, South, East and West approaches respectively. Assuming the framework is employed for $T$ time period, it gathers data from PT every $s$ time period, to observe the state of PT, such as number of vehicles in the scenario, waiting time of each inbound lanes and so on. The specific objective of this framework is to minimize the average waiting time of vehicles as stated in equation 2.

Equation 2 represents average accumulated waiting time of all the vehicles belong to inbound lanes, where $WT_{v_{i}}$ is the waiting time of vehicle ${v_{i}}$ which belongs to lane $h$ and $n_{h}$ represents the number of vehicles on lane $h$ at time $t$. To achieve the objective, this framework disseminates ‘yield’ or ‘wait’ action messages to the CAVs according to Algorithm 1. The algorithm utilizes a function $wt$ to determine the cumulative waiting time of the vehicles on an inbound lane. The framework quantifies the waiting time of a vehicle using the duration of a vehicle’s stationary state [10]. To accomplish this, it utilizes speed-data of vehicles acquired from the DT simulation. After calculating the accumulated waiting time for each inbound lane, various strategies involving different lane combinations i.e. $H$, $y_{2}$ and $y_{3}$ are simulated in the DT instances to facilitate CAV passage. From these simulations, the average waiting time $sim_{wt}$ for each strategy over the next $p$ time period is extracted. The strategy yielding the minimum $sim_{wt}$ value is labeled as $y^{*}$, and the AVs belong to set $V_{y^{*}}$ which is associated with strategy $y^{*}$, are instructed to yield in the PT roundabout. To identify $y^{*}$, this framework prioritizes lanes based on their accumulated waiting time and comes up with 3 strategies utilizing specified combinations of lanes as outlined below:

III-C Flow-Table Space Optimization for Vehicular SDN

Short-lived BSM data packets are managed reactively in SDN to enable bidirectional communication between applications and CVs. An increase in CVs can cause flow-table flooding due to limited TCAM capacity, resulting in delays and reduced throughput [11]. Premature release of flow entries may lead to frequent re-installations and processing delays. To address this, we optimize flow entry and release policies to minimize overflow and re-installations. We simulate three early-release strategies: removing inactive entries based on flow statistics or idle timeouts [7], using a hard timeout for entry removal, and employing an early-release technique for UDP flows to maintain continuous data exchange between CVs and edge-hosted applications.

In vehicular-SDN, a packet_in event triggers the switch to request a flow setup from the controller if it lacks a rule for an incoming packet from a $v_{CV}$. The controller can either program a rule to direct packets from $v_{CV}$ to the application, use the packet_out mechanism to route packets through a specific port, or drop them. Properly configuring a rule or flow entry for $v_{CV}$ is crucial to avoid overloading the controller with packet_out operations, which could disrupt network operations. Since applications need data from $v_{CV}$ to send instructions, each $v_{CV}$ requires 2 flow entries. Each SDN switch has a limited number of flow entries, and different edge applications require separate entries. Assuming a maximum of $f_{max}$ flow entries per application, flow-table overflow occurs if the number of CVs using all applications exceeds half of $f_{max}$. At time $t$, if $V_{CV}$ represents the set of all $v_{CV}$, the number of overflows $N_{of}$ can be determined as shown in equation 3.

Assuming the set of installed flows $F_{t}$ in a switch and set of removed flows $R_{t}$ from the same switch at time $t$, the number of re-installed flows, $N_{re}$ at that time can be calculated using equation 4.

On top of that, as the UDP flows do not ask for connection termination, it is challenging to release the flow entries when they become inactive. So it is necessary to maintain a low table occupancy to ensure overall network operations. To measure the maximum flow entry occupancy for each application, this work implements equation 5, which determines the maximum value of the ratio between $F_{t}$ and $f_{max}$ for the time ranging from $1\text{ to }T$, where $T$ is the current timestamp.

This system seeks to accomplish the objective in equation 6, by alleviating $P(of)$, the probability of flow-table overflow and $P(re)$, the probability re-installation of flow entries, while simultaneously reducing $R_{oc}$ and ensuring the existence of essential flow entries to facilitate communication between the application and CVs approaching the inbound lanes.

IV-A Implementation and Experimental Settings

To develop and evaluate the proposed DT, this study implements the functionalities and conducts unit testing using prerecorded videos of the Pullen Road Roundabout at Pullen Road – Hillsborough Street. As depicted in Fig. 4, these videos are recorded using an unmanned arial vehicle (UAV) and are fed to the DataFromSky-TrafficSurvey system to simulate real-time video capturing and processing of the DT. Essential information is extracted from the system output to create the vehicle-object, as discussed in the previous section. To replicate the roundabout scenario and conduct traffic simulations, we use OpenStreetMap (OSM) to create a 4-way roundabout road network, excluding pedestrian crossings, based on the Simulation of Urban MObility (SUMO). In the simulation, a variable penetration rate is applied to randomly choose a vehicle and simulate AV with the ACC car-following model. This allows better realization of the proposed data driven applications’ scope, the impact of V2I-based connectivity, and to enhance experimental flexibility. To simulate SDN, flow-table, and manage flow-table entries, we use our custom module alongside the Mininet network emulator and Ryu SDN controller. To verify the accuracy of traffic flow simulation by the DT, we measure and analyze the traffic flow from both the PT and DT according to the method in Fig. 4. In Fig. 5, it is evident that due to the AV car-following model, a discrepancy exists between the PT data and the DT simulation data, which can be minimized by optimizing the AV model. For simulation, this study assumes that the total set of vehicles in the scenario consists of HV and AV, where only AVs are connected to the framework through V2I links. Given the penetration ratio of AVs and traffic volume, the number of HVs is calculated using equation 7, where $Vol_{HV}$ denotes the volume of HVs, $Vol_{total}$ signifies the total number of vehicles per hour in the scenario, and $r_{AV}$ represents the penetration rate of AVs.

To assess the flow management system’s efficiency in utilizing flow table space, various $f_{max}$ values are used to simulate flow entry management scenarios, the 3 strategies, idle timeout, random hard timeout, and the proposed strategy are comparedwith no-timeout simulation. Timeout values can be set with flow entries to ensure they are released when the timeout expires. In this study, for random hard timeout, a random value between 10 and 300 seconds is used. In case of proposed method, the operating distance is set to 50 meters. Different levels of AV penetration rates with V2I based connectivity are also incorporated into these experiments to simulate the natural progression of flow entry over time.

IV-B Result Discussion and Performance Evaluation

The simulation of the decision support framework demonstrates that increasing the penetration rate of AVs, alongside the traffic flow of HVs, reduces average waiting time. However, once the AV penetration rate exceeds 60%, the waiting time begins to rise again. Fig. 6 shows the relationship between waiting time and penetration rate of AV, aligning with findings in [12]. Nonetheless, the simulation results indicate that average waiting time can be decreased through connectivity between AVs and infrastructure, enabling control from the decision support framework. The EC-based decision support framework, as illustrated, improves traffic efficiency with the increasing penetration rate of AVs, achieving a 22% reduction in waiting time compared to only HVs, and an additional 18% reduction in waiting time compared to HVs and AVs, even with a modest 40% penetration rate of AVs.

As illustrated in Fig. 7, the probability of overflow increases with the decreasing number of $f_{max}$. However, if no timeout is set for the flows, overflow inevitably occurs, even when $f_{max}$ reaches 250.In contrast, when the proposed strategy is implemented, overflow does not occur, even with $f_{max}$ set to 125 and a 100% penetration rate of CAVs. Flow re-installation is influenced more by the flow entry duration or timeout value than by $f_{max}$, as previously noted. A no-timeout approach prevents flow re-installation since flows are not released. In contrast, a random hard timeout value increases re-installation probability, as CVs continue to transmit packets post-timeout. Implementing an idle timeout minimizes re-installation chances by allowing flow entries to persist longer. The proposed method, with a timeout duration that accommodates vehicle traversal, also shows a low re-installation probability. As shown in Fig. 8, $R_{oc}$ decreases with increasing $f_{max}$ in terms of flow table occupancy.Yet, with no timeout, occupancy is nearly always at its maximum, even with $f_{max}$ set to 500. In this scenario, a random hard timeout performs better than an idle timeout because not all the flows wait for the CVs to stop BSM transmission. Nonetheless, the proposed strategy shows significantly lower occupancy, using only 50% of the flow table even with a 100% penetration rate and $f_{max}$ set to 250, which is roughly half the occupancy compared to other strategies. Fig. 9 illustrates the value of $y$ across various strategies. It indicates that, compared to alternative methods, the proposed strategy decreases chances of table overflow, re-installation, and reduces flow-table occupancy, even under constraints on flow-table size allocation. Moreover, it ensures the creation of flow entries to facilitate effective communication between the applications and CVs.