Digital Twin-Assisted Cooperative Driving at Non-Signalized Intersections

A. Background B

• Intersections are traffic conflict locations, and reported crash data and driver-attributed causes motivate attention to intersection safety.A cited U.S. Department of Transportation report records 2.4 million intersection-related crashes in 2007, 40% of reported crashes and 21.5% of traffic fatalities.Source: A. Background · p. 1

B. Literature Review

• Prior approaches include reservation-based intersection management, virtual-platoon motion control, optimization methods, and real-world or mixed-reality implementations.Some optimization formulations can impose high computational loads that limit real-time implementation.Source: B. Literature Review · pp. 1–2

C. Contributions of this Study

• Existing work often assumes fully connected automated vehicles, leaving little guidance for human drivers cooperating with CAVs, and commonly assumes perfect V2X communication despite delay and packet loss.Source: C. Contributions of this Study · p. 2
• The study combines an enhanced FIFO slot reservation algorithm, consensus motion control, and model-based motion estimation with an AR HMI for human-driven connected vehicles.The HMI visualizes vehicle Digital Twins as reserved slots to guide cooperative driving.Source: C. Contributions of this Study · p. 2

D. Organization of this Study

• The study is organized from problem statement and system architecture through cooperative-driving methods, AR HMI, simulation and evaluation, then conclusions and future directions.Source: D. Organization of this Study · p. 2

II. PROBLEM STATEMENT AND SYSTEM ARCHITECTURE

• The digital world has motion and communication modules; motion components include map matching, path planning, FIFO slot reservation, and consensus control, while estimated target-vehicle motions feed motion control.Source: §II · p. 3
• Physical connected vehicles include localization, motion measurement, gateway, optional perception, and a Digital Twin-assisted AR HMI for visualized driver assistance.Source: §II · p. 3

• Map matching synchronizes GNSS vehicle coordinates with a pre-built map and applies geo-fencing to determine whether downstream motion-module components should operate.The map stores intersection geometry, road and lane attributes, waypoints, direction, and speed limits.Source: §III · p. 3
• Path planning generates lane-level intersection routes from vehicle origins and destinations, enabling conflict-point and target-vehicle identification and distance calculations.For a four-leg, single-lane intersection, the passage identifies 16 crossing, 8 merging, and 8 diverging conflicts.Source: §III · p. 4
• Map matching and path planning are implemented in the simulation, but their technical details are not covered in the study.Source: §III · p. 4

A. Enhanced FIFO Slot Reservation

• Unlike FIFO or FCFS rules triggered simply by geo-fence entry, the enhanced FIFO algorithm also considers estimated time of arrival to a conflict point.ETA is based on an assumed intersection speed limit and the vehicle's preferred acceleration.Source: A. Enhanced FIFO Slot Reservation · p. 4
• Approach ETA distinguishes cruising at the speed limit from lower-speed cases, where the vehicle either cannot or can accelerate to the speed limit before the conflict point.The two lower-speed cases are illustrated as distinct acceleration profiles.Source: A. Enhanced FIFO Slot Reservation · p. 4, Figure 2 · p. 4
• For vehicles following a preceding vehicle, ETA is updated to maintain a constant car-following time headway, so the follower's ETA cannot precede the leader's.Source: A. Enhanced FIFO Slot Reservation · pp. 4–5
• ETA-aware, location-aware reservation prevents a slower vehicle entering a geo-fence first from forcing a conflicting vehicle to decelerate and reducing throughput and energy efficiency.A reservation request is triggered either when ETA falls below a predefined time constant or when the vehicle enters the intersection geo-fence.Source: A. Enhanced FIFO Slot Reservation · p. 5

B. Consensus Motion Control

• The controller takes assigned slots and target vehicles as inputs and computes the ego vehicle’s referenced motion.The digital-world motion-control component focuses on high-level control that generates the reference motion.Source: B. Consensus Motion Control · p. 5
• The vehicle-dynamics model relates longitudinal motion to engine force, braking torque, aerodynamic drag, friction, and mechanical drag.The section’s high-level controller generates references; low-level control commands actuators to track them.Source: B. Consensus Motion Control · p. 5
• A double-integrator consensus controller generates longitudinal motion using desired headway and tuned control gains.The gains are tuned through feedforward control to support the safety, efficiency, and comfort of slot-following.Source: B. Consensus Motion Control · p. 5
• A lookup table dynamically sets the control gains from the two vehicles’ initial speeds and initial headway.The control gains are k_ij and γ_i.Source: B. Consensus Motion Control · p. 5
• The controller drives vehicle i’s speed toward vehicle j’s speed and its longitudinal distance toward j’s, with a distance difference retained.This is the stated convergence behavior of the longitudinal motion-control algorithm.Source: B. Consensus Motion Control · p. 6
• Communication delay makes received target-vehicle information stale, so a model-based estimator adjusts the referenced motion before physical execution.The delayed-control formulation uses target-vehicle information from t − τ_ij(t).Source: B. Consensus Motion Control · p. 6

C. Model-Based Motion Estimation for Communication Delay and Packet Loss

• The proposed motion estimator tackles communication delay and packet loss together, enabling cooperative driving with moderate levels of both impairments.The authors contrast this joint treatment with prior work that addressed the issues separately,.Source: C. Model-Based Motion Estimation for… · p. 6
• The main estimation function calls three subalgorithms to produce vehicles’ future longitudinal motion estimates.The output is estimated longitudinal speed and distance over the future horizon [t + 1, t + N].Source: C. Model-Based Motion Estimation for… · p. 6
• For a vehicle without a target to follow, the IDM-based estimate predicts future speed converging to its known, preset target speed.The algorithm describes the vehicle as a leader in the communication topology.Source: C. Model-Based Motion Estimation for… · p. 6
• During packet loss, the future speed estimate remains unchanged when no information update arrives; exceeding the safety threshold triggers an all-way-stop fail-safe.The fail-safe transfers the upcoming non-signalized intersection into an all-way stop.Source: C. Model-Based Motion Estimation for… · p. 6
• Future longitudinal distance is estimated cumulatively from the predicted speed trajectory.The update adds the preceding predicted speed multiplied by the prediction-step duration.Source: C. Model-Based Motion Estimation for… · p. 6
• For a following vehicle, the speed estimate accounts for communication delay, treating the target’s speed differently depending on whether delay is shorter than one prediction step.The algorithm also adjusts the target vehicle’s estimated longitudinal distance for delay.Source: C. Model-Based Motion Estimation for… · p. 7

IV. DIGITAL TWIN-ASSISTED AR HMI

• The HMI is an optional design for human-driven connected vehicles, and driver compliance is outside this study’s scope.Source: §IV · p. 7
• Reserved red slots represent target vehicles from other intersection legs, while the driver is guided to travel within green available slots.The HMI is optional, and user compliance is outside the study’s scope.Source: §IV · p. 7, Figure 4 · p. 7
• Slot length is adjusted for the following clearance, and its redundancy factor can be tuned to V2X performance or motion-estimation confidence.A higher factor means more redundancy.Source: §IV · p. 7
• A pinhole-camera transformation maps slot coordinates from the world frame to the AR image plane for display to the driver.The transformation uses extrinsic parameters and projection-device parameters such as focal length and lens distortion.Source: §IV · p. 8

A. Modeling and Simulation Environment in Unity Game Engine

• The Unity environment uses a SoMa-based corridor with four intersections and centimeter-level routes along five streets to support map matching and path planning.The system algorithms are modeled for each connected-vehicle agent, and the AR HMI is modeled on the ego vehicle.Source: A. Modeling and Simulation Environment… · p. 8, Figure 6 · p. 9
• Communication delay is modeled from an LTE-based V2X field test as a time-variant normal distribution with a 40 ms mean and 0.0259 standard deviation.The authors use LTE, which they describe as normally having higher delays than DSRC, as a stress test.Source: A. Modeling and Simulation Environment… · p. 8
• Packet loss combines random losses with losses during selected periods to represent periodic issues and NLOS obstruction, at a frequency above the expected real-world rate.The elevated frequency is used to stress-test motion estimation.Source: A. Modeling and Simulation Environment… · p. 8
• HITL testing uses a Windows desktop, Logitech G29 wheel, and Unity 2019.2.11f1 driving simulator.The computer uses an Intel Core i7-9750 processor, 32 GB memory, and NVIDIA Quadro RTX 5000 Max-Q graphics.Source: A. Modeling and Simulation Environment… · p. 9, Figure 7 · p. 9
• Participants drive from Bryant Street through four consecutive non-signalized intersections, then repeat the route with fixed-timing signals as the baseline.In the baseline, NPC vehicles are randomly generated and have no CAV features; in the non-signalized scenario, they run the proposed algorithms as CAVs.Source: A. Modeling and Simulation Environment… · p. 9

B. Human-in-the-Loop Simulation Results

• In a human-driven ego-vehicle example, the ego maintains relatively safe distances while three NPC vehicles consecutively decelerate from 5–12 s to maintain safe following distances.The NPC vehicles use the proposed motion-control and motion-estimation algorithms.Source: B. Human-in-the-Loop Simulation Results · p. 9, Figure 9 · p. 10
• Slot reservation operates independently at each intersection, continuing as vehicles approach and leave; the ego receives its next slot before two NPCs clear the first intersection.Source: B. Human-in-the-Loop Simulation Results · p. 9
• In the sample four-intersection comparison, the proposed scenario maintains relatively stable speed without full stops, unlike the signalized baseline’s red-light stops at the first and fourth intersections.The baseline reaches a higher maximum speed, but its excessive speed changes increase travel time and energy consumption.Source: B. Human-in-the-Loop Simulation Results · pp. 9–10, Figure 10 · p. 10
• 20% average travel-time reduction and 23.7% average fuel-consumption reduction were achieved versus the baseline signalized scenario.Fuel consumption was calculated with the open-source MOVESTAR model, assuming gasoline vehicles.Source: B. Human-in-the-Loop Simulation Results · p. 10

C. Sensitivity Analysis on Motion Estimation Algorithm

• The analysis measures each trip’s maximum longitudinal position-estimation error and minimum simulation frame rate under time-varying delay and hybrid packet loss.Source: C. Sensitivity Analysis on Motion… · p. 10
• At a 0.01 s prediction time step, maximum position estimation error stays below 0.2 m; at 1 s, it can reach 5.8 m during major packet loss.The authors attribute the difference to information granularity and update frequency.Source: C. Sensitivity Analysis on Motion… · p. 10
• At a 0.01 s prediction time step, minimum simulation frame rate drops to 6 FPS; at 1 s, it remains above 19 FPS.The authors attribute the load difference to how many times motion-estimation Algorithm 2 is called.Source: C. Sensitivity Analysis on Motion… · p. 10
• Because multiple vehicles are simulated centrally on one computer, cloud or distributed computing may relieve the computational-load concern; balancing accuracy and load remains future work.Source: C. Sensitivity Analysis on Motion… · pp. 10–11

VI. CONCLUSION AND FUTURE WORK

• The Digital Twin framework combines enhanced FIFO slot reservation, consensus motion control, model-based motion estimation, and AR HMI guidance for connected vehicles crossing without full stops.The algorithms schedule crossing order, generate referenced vehicle motion, and address V2X communication issues.Source: §VI · p. 11
• 20% reduction in travel time and 23.7% reduction in energy consumption were reported versus traditional signalized intersections in HITL simulation.Agent-based simulation used Unity and a real-world San Francisco map; sensitivity analysis examined prediction-timestep effects on accuracy and computational load.Source: §VI · p. 11
• The modeled system does not yet cover mixed traffic involving legacy vehicles and vulnerable road users such as pedestrians and bicycles.The paper identifies cooperative driving with connected vehicles, legacy vehicles, bicycles, and pedestrians together as an open question.Source: §VI · p. 11
• Future work proposes using on-board perception data as an additional motion-estimation source and comparing the method with other intersection-control approaches.The suggested comparisons include Eco-Approach and Departure and dynamic or adaptive signal control with connected vehicles.Source: §VI · p. 11