Beacon: Traffic Blackout Dataset Benchmark

• Beacon is a comprehensive traffic dataset that captures naturalistic driving behavior during urban signal failures, providing lane-level vehicle observations from Memphis intersections.
• It employs a SUMO-based reconstruction pipeline to analyze unsignalized, signalized, and mixed-traffic regimes with detailed mismatch metrics across different blackout scenarios.
• Beacon serves as a benchmark for assessing control strategies, evaluating the impact of robot vehicles on wait times, travel times, and CO2 emissions under varied demand conditions.

Beacon is a naturalistic driving dataset and benchmarking framework for traffic reconstruction and control during traffic-light blackouts at complex urban intersections. It comprises four hours of vehicle-level observations from two adjacent intersections on Walnut Grove Road in Memphis, Tennessee, captured during an actual storm-related outage, and records for each vehicle a discrete timestep together with its origin and destination lanes. The dataset is explicitly designed to support microscopic reconstruction, density analysis, and evaluation of unsignalized, signalized, and mixed-traffic control regimes under a failure mode that is operationally important but empirically underdocumented. The authors argue that Beacon is likely the first publicly available traffic dataset focused on naturalistic driving behavior at complex intersections during traffic-light blackouts (Sarker et al., 2024).

1. Definition and research motivation

Beacon was introduced to address a specific empirical gap: the absence of publicly available, real-world traffic data describing how drivers behave at complex urban intersections when signal control fails during a blackout. In the motivating account, intersections are critical nodes of traffic networks and account for more than 45% of U.S. crashes. Because modern traffic lights are fully dependent on electricity, blackouts disable signal control and force drivers to self-organize through informal right-of-way negotiation, gap acceptance, and courtesy, often producing congestion, safety risks, and gridlock (Sarker et al., 2024).

The dataset is therefore both an observational resource and a benchmark. As a dataset, it records traffic during the blackout-induced unsignalized state itself. As a benchmark, it provides the demand and lane-transition structure needed to reconstruct that state in SUMO and then compare it with hypothetical signalized and mixed-autonomy regimes. This dual role is central to the project: Beacon is not merely a collection of counts, but a structured basis for replay, inference, and control experimentation.

A further motivation is methodological. Much existing traffic data is collected under normal signalized operation, on highway segments, at simpler junctions, or in heavily instrumented environments such as UAV campaigns, V2X corridors, or proving grounds. Beacon instead documents naturally occurring behavior under a rare but operationally consequential infrastructure failure. This suggests a shift from evaluating control only under nominal conditions to evaluating it under degraded urban operating modes.

2. Dataset composition and vehicle-level representation

Beacon covers two intersections in Memphis, Tennessee: Walnut Grove–Goodlett Street (GWG; latitude 35.131508, longitude -89.925530) and Walnut Grove–Mendenhall Road (WGM; latitude 35.130825, longitude -89.898503). Each intersection was recorded for one midday hour and one afternoon-peak hour on a weekday, yielding four one-hour scenes: GWG-N (12:00–1:00 PM), GWG-AN (5:00–6:00 PM), WGM-N (12:27–1:27 PM), and WGM-AN (5:21–6:21 PM). All four scenes correspond to blackout-induced unsignalized operation (Sarker et al., 2024).

The raw overhead videos were manually annotated by one primary annotator and validated by two additional annotators; the annotators were 25–35 years old. Beacon contains 8,811 distinct vehicles across the four scenes: 1,983 in GWG-N, 2,453 in GWG-AN, 2,033 in WGM-N, and 2,342 in WGM-AN. The highest-demand scene is GWG-AN, and the lowest-demand scene is GWG-N.

For each vehicle $i$, Beacon stores a tuple

$$v_i = \big(t_i,\; l^{\text{orig}}_i,\; l^{\text{dest}}_i\big),$$

where $t_i \in \mathbb{Z}$ is the discrete timestep at which the vehicle reaches the head of its start lane, and $$l^{\text{orig}}_i, l^{\text{dest}}_i \in \mathcal{L}$$ are the origin and destination lane identifiers. This representation is lane-level and vehicle-level, but not continuous in physical coordinates: Beacon does not directly provide full trajectories $x_i(t), y_i(t)$. Instead, it provides the structural information needed to reconstruct those trajectories in a microscopic simulator.

The temporal resolution is aligned with SUMO simulation steps, while density analysis in the paper is conducted at per-minute granularity. Spatially, the representation is lane-specific, with numbered approach and exit lanes. This choice makes the dataset compact but still sufficiently expressive for route inference, flow estimation, and simulator-based replay.

3. Blackout traffic behavior captured by the data

Because the outage was not staged and drivers were not instructed, Beacon captures naturalistic responses to an unexpected loss of signal control. The scenes show heterogeneous behaviors, including cautious and aggressive driving, varied turning preferences, and self-organized right-of-way negotiation. The dataset also exhibits systematic differences between midday and afternoon peak traffic (Sarker et al., 2024).

Demand is strongly shaped by the arterial role of Walnut Grove Road. At GWG, eastbound flows dominate in both midday and afternoon scenes; at WGM, westbound flows dominate. The paper formalizes directional demand as

$Q_d = \sum_{i:\,\mathrm{dir}(v_i)=d} 1,$

and gives hourly directional totals. For example, GWG-AN contains 425 northbound, 629 southbound, 794 eastbound, and 605 westbound vehicles, for a total of 2,453.

Turning behavior is also intersection-specific. At GWG, northbound right turns dominate left turns, while southbound left turns exceed right turns. At WGM, northbound left turns significantly surpass right turns. These asymmetries are visible in lane-transition matrices

$$T_{ij} = \text{number of vehicles with start lane } i \text{ and end lane } j,$$

which encode the dominant through movements and turning distributions for each scene.

Per-minute traffic density is defined as

$$\rho(t) = \text{number of vehicles observed during minute } t,\qquad t=1,\dots,60.$$

The reported densities are approximately 25–35 vehicles per minute in GWG-N, 35–45 in GWG-AN, 35–45 and relatively stable in WGM-N, and 35–42 with high fluctuation in WGM-AN. The authors attribute the instability of WGM-AN to higher demand, more complex lane geometry and local network structure, and variation in driver caution, assertiveness, and trip purpose. A plausible implication is that blackout instability is not a uniform property of unsignalized control, but depends materially on intersection geometry and directional demand balance.

4. Reconstruction pipeline and benchmark scenarios

Beacon is explicitly designed for microscopic reconstruction in SUMO. The reconstruction workflow begins by extracting the physical road network from OpenStreetMap and using NETCONVERT to create a SUMO .net.xml representation that preserves lane counts, geometry, and intersection layout. Integer lane IDs in Beacon are then mapped to SUMO lane identifiers. For each vehicle tuple $v_i = (t_i, l^{\text{orig}}_i, l^{\text{dest}}_i)$, the simulator inserts the vehicle at time $t_i$ into the corresponding origin lane and assigns a route

$$v_i = \big(t_i,\; l^{\text{orig}}_i,\; l^{\text{dest}}_i\big),$$0

consistent with the recorded entry and exit lanes. Vehicle insertion and control are implemented through TraCI (Sarker et al., 2024).

The paper evaluates three reconstruction regimes. The first is the original unsignalized blackout state. The second overlays hypothetical fixed-time traffic signals on the same demand. The third introduces mixed traffic with robot vehicles and human-driven vehicles. These derived scenarios convert Beacon from a passive corpus into a benchmark for model fidelity and controller comparison.

For unsignalized reconstruction, correctness is assessed by timestep mismatches, start-lane mismatches, and end-lane mismatches. Reported total mismatch rates are 0.41% for GWG-AN, 1.93% for GWG-N, 8.4% for WGM-N, and 8.75% for WGM-AN. The smaller errors at GWG indicate that the simpler intersection geometry is easier to reconstruct faithfully than WGM.

For signalized reconstruction, simulated stopping and queueing break the direct correspondence between simulated and observed timesteps, so evaluation focuses on lane mismatches only. The fixed-phase signal plans are specified directly: for GWG, the cycle durations are 7, 4, 29, 4, 20, 4, 4, 40, and 4 seconds; for WGM, they are 22, 4, 80, 4, 42, 4, 4, 113, and 4 seconds. Start/end lane mismatch rates under these signal plans are 0.612% for GWG-N, 4.49% for GWG-AN, 14.13% for WGM-N, and 15.77% for WGM-AN. This suggests that fixed-time signalization can materially alter exit-lane usage relative to the blackout state, especially at WGM.

5. Mixed traffic control and robot-vehicle experiments

Beacon is also used to study mixed autonomy by labeling vehicles as either human-driven vehicles (HVs) or robot vehicles (RVs). RVs are controlled algorithmically, often through reinforcement-learning-based methods, while HVs follow standard car-following and gap-acceptance models. The RV penetration rate is denoted $$v_i = \big(t_i,\; l^{\text{orig}}_i,\; l^{\text{dest}}_i\big),$$1. Performance is summarized with three metrics: average wait time $$v_i = \big(t_i,\; l^{\text{orig}}_i,\; l^{\text{dest}}_i\big),$$2, average travel time $$v_i = \big(t_i,\; l^{\text{orig}}_i,\; l^{\text{dest}}_i\big),$$3, and average CO$$v_i = \big(t_i,\; l^{\text{orig}}_i,\; l^{\text{dest}}_i\big),$$4 emission rate per vehicle $$v_i = \big(t_i,\; l^{\text{orig}}_i,\; l^{\text{dest}}_i\big),$$5 in mg/s (Sarker et al., 2024).

The empirical pattern is intersection-dependent. At GWG, under current demand, RVs provide limited improvement because human drivers already perform reasonably well. At WGM, which is more complex and more congested, increasing RV penetration reduces both wait times and travel times, particularly in the busier scenes. The effect therefore appears contingent on both geometry and load.

The paper also tests counterfactual demand scaling at GWG. When demand is increased by 25%, RVs reduce wait times by up to 82.6% and travel times by 10.3%, with a 7.2% increase in CO$$v_i = \big(t_i,\; l^{\text{orig}}_i,\; l^{\text{dest}}_i\big),$$6. When demand is increased by 50%, RVs reduce wait times by 47.1% and travel times by 21.8%, again with higher emissions attributed to increased movement and less idling. Within the paper’s framing, these experiments demonstrate that Beacon can benchmark not only reconstruction accuracy but also capacity, robustness, and efficiency under mixed-control interventions.

The benchmark implications are direct. Since Beacon supplies realistic lane-level entries, exits, demand patterns, and density traces, it can support evaluation of unsignalized control strategies, fixed-time and adaptive signal control, and mixed-traffic policies. The paper explicitly notes evaluation via mismatch rates, directional flow differences, turning counts, per-minute density patterns, wait time, travel time, CO$$v_i = \big(t_i,\; l^{\text{orig}}_i,\; l^{\text{dest}}_i\big),$$7, and safety proxies such as conflict frequency, speed variance, and stop-and-go behavior.

6. Novelty, limitations, and nomenclature

The paper summarizes Beacon’s contribution in four parts: a vehicle-level blackout dataset for two Memphis intersections over four hours and 8,811 vehicles; empirical analysis of demand, transitions, turning, and density; an end-to-end OSM-to-SUMO reconstruction pipeline; and comparative control experiments spanning unsignalized, signalized, and mixed-traffic settings (Sarker et al., 2024).

Its claimed novelty is contextual rather than merely volumetric. Beacon is positioned against datasets such as INTERACTION, USyd, and V2X-Seq. Those resources contain large numbers of trajectories or scenarios, but they are not centered on blackout-induced unsignalized control at complex urban intersections. Beacon differs by focusing on a naturally occurring signal-failure regime, representing lane-specific vehicle entries and exits for microscopic reconstruction, and using a low-instrumentation collection setup based on a single overhead camera per scene.

The limitations are equally explicit. Reconstruction fidelity is high at GWG but more challenged at WGM; the dataset covers only two intersections and four hours; larger-scale city networks and additional blackout events remain future work; and emissions and safety are presently handled through proxies rather than more direct microscopic models. The authors also state plans to integrate Beacon with large-scale traffic prediction, adversarial training, and network optimization.

The term “Beacon” is also used in unrelated arXiv literatures, so the traffic dataset should not be conflated with the Ethereum 2.0 Beacon Chain (Cassez et al., 2021), the BEACON early-exit framework for automatic modulation classification (Liu et al., 9 Apr 2026), or Activation Beacon for long-context compression in LLMs (Zhang et al., 2024). In transportation research, however, Beacon most specifically denotes the blackout traffic dataset and benchmark described here: a record of how urban intersections behave when conventional signal control fails, and a platform for reconstructing, comparing, and controlling those behaviors under realistic demand.