arXiv:2609.31429v1 [cs.NI] 25 Sep 2026
ITS Fairy: Occlusion Assistance Selected Against a Recipient’s Own Perception Reports Yenan Wang, Oscar Karlsson, and Elad M. Schiller
Francesco Raviglione∗ and Claudio Casetti†
Department of Computer Science and Engineering Chalmers University of Technology Gothenburg, Sweden {yenan, oscakarl, elad}@chalmers.se
∗ Department of Electronics and Telecommunications
Abstract—Cooperative perception can expose object state beyond a vehicle’s onboard sensors, but sensing occlusion can still leave a local safety application without the objects its collision computation needs. To tackle this challenge, we present the ITS Fairy, an infrastructure-side Server Local Dynamic Map (SLDM) service whose decision unit is the pair (recipient, missing conflict-relevant object): among objects absent from a recipient’s CPM-derived reported awareness, it sends only those relevant to a Time of Closest Approach (TCA) conflict test. Comparable services predict what a vehicle can perceive; the ITS Fairy instead reads what it has already reported. The recipient inserts the selected state into its local LDM and uses its unchanged collisionavoidance controller. We evaluate this application-level mechanism in SUMO–ms-van3t–S-LDM emulation, since extended as VaN3Twin, using a sensing-occluded lane merge and four-way intersection scenario. At every main-sweep speed, the smallest assisted per-encounter minimum TCA exceeds the largest localonly value in the archived data. Additionally, assisted medians remain in the multi-second range where local-only operation repeatedly approaches zero. In the lane-merge robustness data, the median benefit persists at 80% configured assistance omission with 10 and 5 Hz analysis, but largely disappears at 100–120 km/h when 80% omission is combined with 1 Hz analysis. These results demonstrate the application-level value of supplying object state selected against what a recipient has itself reported. They are not a vehicular wireless-channel evaluation, and they do not quantify what selectivity saves relative to forwarding every nearby object. Index Terms—cooperative perception, infrastructure-assisted perception, server local dynamic map, sensing occlusion, collision avoidance
I. I NTRODUCTION A collision-avoidance application can act only on the traffic participants represented in its local environment model. Sensing occlusion can hide exactly the object that a conflict computation needs: two vehicles approaching an obstructed junction may be on intersecting paths while neither appears in the other’s Local Dynamic Map (LDM). Cooperative perception and an infrastructure-hosted Server Local Dynamic Map (S-LDM) [1] may nonetheless hold that object state, aggregated from the CAMs and CPMs that multiple vehicles already transmit [2], [3]. Providing all of it to every vehicle is wasteful of the shared wireless channel, so a recipient-specific question arises: which part should be supplied to this vehicle? Open Issues and Positioning. Infrastructure-assisted perception and its operational constraints are already well studied, in-
† Department of Control and Computer Engineering
Politecnico di Torino Turin, Italy {francesco.raviglione, claudio.casetti}@polito.it
cluding end-to-end delay in cloud-assisted pipelines [4], roadside sensor fusion [5], communication delay in context-aware collision avoidance [6], information freshness [7], receiverside reduction of redundant CPM processing [8], and a real obstructed intersection with active traffic-moderating infrastructure [9]. What remains open is not whether infrastructure can assist, but how the information it provides is selected. The simplest selection rules consult only the sender’s own object list and the state of the channel. The Collective Perception Service (CPS) [3] filters through generation rules and redundancy mitigation applied to the sender’s own detections; RRS [10] ranks objects by an accident-risk indicator and by how often each has already been reported; and AICP [11] sorts objects by informativeness to limit channel load and on-screen clutter. These rules establish what is worth sending in general. However, none of them determines what a particular vehicle is missing to enable collision avoidance effectively. Another family of rules relies on predicted risk rather than on the recipient’s perceptual state. Ma et al. [12] decide whether cooperation is warranted for a road user, from collision risk and blind-spot occupancy; SRA-CP [13] has vehicles exchange coverage summaries and select peers for risk-relevant blind zones; and STDR-CP [14] aggregates pervehicle demand into a global broadcast schedule. The closest related work, IBCP [15], conditions on an ego’s declared intent and returns conflict alerts aimed only at the road users relevant to that ego’s announced maneuver, evaluated with a minimummargin metric under occlusion. Risk and intent identify which encounters matter, but not which participant lacks the object state that makes the encounter dangerous. Closer to this work are approaches that estimate what a particular recipient already holds. AutoCast [16] computes, for every object and neighbour, whether the object is geometrically visible from that neighbour’s pose and whether it is relevant to the neighbour’s broadcast trajectory. CCPAV [17] scores an object for a specific receiver as its unseen-ness divided by the time until the two trajectories intersect, and arbitrates centrally under a resource constraint. Lv et al. [18] learn perceptibility from object geometry and occlusion, and disseminate only the objects predicted to be unseen. Higuchi et al. [19] maintain, for each neighbour, an anticipated history of the messages it is likely to have received, and include
TABLE I W ORK , AWARENESS SOURCE , AND RESULTING ACTION . Work
How awareness is obtained
Resulting action
RRS [10]
Not modelled; object redun- Fixed-size object set in a broaddancy counted across received cast CPM. CPMs. AICP [11] Not modelled; objects sorted by Filters objects for forwarding informativeness. and driver display. Ma et al. [12] Not modelled; collision risk and Identifies CP necessity from inblind-spot occupancy. frastructure observations. IBCP [15] The ego’s declared intent. Issues conflict alerts into a dedicated braking pathway. Lv et al. [18] Estimated: learned perceptibil- Disseminates objects predicted ity from object geometry. imperceptible. AutoCast [16] Estimated: geometric visibility Schedules point-cloud transmisfrom the neighbour’s pose. sion between vehicles. CCPAV [17] Estimated: receiver line- Prioritises centrally relayed of-sight, scaled by time to messages. trajectory intersection. Higuchi et al. [19] Estimated: anticipated message Includes records that change an history per neighbour. inferred belief. Abdel-Aziz et al. [20] Estimated: learned from the Transmits selected regions to paired receiver’s feedback. the paired vehicle. ITS Fairy Read: the recipient’s own re- Sends object state into the recipported CPM history. ient LDM; maneuver remains local.
a perception record only when it would measurably change that neighbour’s inferred belief. Abdel-Aziz et al. [20] pair vehicles at the infrastructure and let each sender learn which regions to share from its partner’s feedback. In all of them the recipient’s awareness is inferred rather than observed, and each estimator carries a cost: geometric visibility presupposes a model of the recipient’s sensors, learned perceptibility is weakest where several vehicles converge, and feedback-driven schemes require a return channel that Higuchi et al. [19] judged too expensive in vehicle-to-vehicle operation to justify the savings. A parallel line of work makes related decisions on sparse feature or point-based representations inside a multiagent model [21], [22] rather than on object state. Table I summarizes this progression, from no recipient model, through estimated recipient state, to what the recipient has reported. The open issue is therefore this: recipientspecific supplementation needs to rest on evidence of what the recipient has reported perceiving, and to send only those missing objects that bear on an actual conflict. Proposed Approach. We propose a novel design concept, named ITS Fairy, whose decision unit is the pair (recipient, missing conflict-relevant object). For each vehicle, the service derives a recipient-specific reported-awareness set from that vehicle’s own CPM history, and forms the residual between the fresh S-LDM objects in the vehicle’s neighbourhood and the objects the vehicle has itself reported. A conflict test based on Time of Closest Approach (TCA), an established criticality metric for automated driving [23], is applied to that residual alone; for each surviving object the Fairy transmits that object’s state, and neither a collision verdict nor a maneuver command. The recipient inserts the supplied state into its local LDM, recomputes TCA, and acts through its unchanged avoidance controller. Westhofen et al. refer to this quantity as Time To Closest Encounter; we use the term TCA throughout.
This is affordable for architectural rather than algorithmic reasons: an S-LDM already ingests the CPMs of every vehicle in its coverage area, so the awareness set is a byproduct of aggregation and requires no additional message, which is why the option dismissed as too costly between vehicles is available here. Holding the local controller unchanged is likewise deliberate, so that an observed difference is attributable to information availability rather than to a different predictor. Evaluation. Two questions drive our evaluation. RQ1: under sensing occlusion, how does recipient-specific missing-object assistance change the minimum-TCA margin available to an unchanged local collision-avoidance pipeline, relative to local-only operation? This is examined in a sensing-occluded lane merge and four-way intersection, with local-only and assisted operation sharing the same local controller, using perencounter minimum TCA as the primary metric and route traversal time as secondary. RQ2: in the lane-merge, how does the retained minimum-TCA benefit vary with configured assistance-message omission and with the Fairy’s analysis rate? Omission is swept from 0 to 80%, and the Fairy’s analysis rate, i.e., how often it recomputes a vehicle’s missing-object residual and tests it for conflict, is set to 10, 5, and 1 Hz. That delivery loss, staleness and interruption matter is established rather than claimed here: object state has been delivered into an unmodified automated-driving stack on real vehicles at an occluded intersection [24], and the effects of unreliable and interrupted communication studied directly [25], [26]. RQ2 asks the narrower question of how this mechanism degrades, not how a wireless channel behaves. Findings. We claim that supplying conflict-relevant object state, selected against what a recipient has itself reported, improves the information available to an unchanged local safety controller under the tested occlusion conditions. Effectiveness. Assistance separates local-only and Fairyassisted operation completely in the obtained results: at every evaluated speed in both geometries, the smallest assisted perencounter minimum TCA exceeds the largest local-only value. Assisted medians hold at multi-second values where local-only medians sit at or below 0.08 s over most of the sweep, and the four-way intersection reproduces the effect in a multi-vehicle geometry. Lane-merge traversal time also falls, by about 11% to 45–47% across 30–120 km/h. Robustness. The benefit is jointly conditioned on omission and analysis frequency rather than on omission alone. At 10 Hz the benefit survives 80% omission almost intact; at 1 Hz the same omission leaves it only at the lowest tested speed, and the collapse is abrupt rather than gradual. Contributions. The ITS Fairy addresses a challenge created by sensing occlusion, which can leave a vehicle’s local safety application without the state of a conflict-relevant object that is nevertheless available at the infrastructure. The Fairy uses the recipient’s own CPM-derived reported awareness to identify such missing objects, filters them for conflict relevance, and supplies their state while collision assessment and maneuver selection remain local. Our contributions are:
Occluded object
CAM/CPM to infrastructure
CAM/CPM reports S-LDM + ITS Fairy
Recipient local LDM selected missing object state
local TCA + maneuver blocked sensing LOS
Fig. 1. ITS Fairy information flow. The S-LDM aggregates CAM/CPM information, while the Fairy selectively supplies object state missing from a recipient’s reported awareness. Collision-risk recomputation and maneuver control remain local.
A novel infrastructure-side awareness model that reads recipient awareness from the recipient’s own CPM reports rather than inferring it, and uses it to identify and supply conflict-relevant missing object state. • An open-source SUMO–ms-van3t–S-LDM evaluation environment implementing the ITS Fairy and the sensingoccluded lane-merge and four-way-intersection scenarios. • An evaluation showing substantial minimum-TCA benefits, and a joint characterization of how they depend on assistance omission and analysis rate. We pledge to release the evaluation environment and the archived result data upon acceptance; Section III states which experimental configurations are preserved in the repository. We expect the ITS Fairy concept, its recipient-specific awarenessrepair method, and the accompanying evaluation environment to provide an open basis for studying targeted infrastructure assistance and to inform future cooperative-perception designs. •
II. S YSTEM M ODEL AND ITS FAIRY System Model. Vehicles maintain local LDMs and generate CAMs describing their state and CPMs describing perceived objects [2], [3]. An infrastructure node receives these reports, maintains a centralized S-LDM, and hosts the Fairy [1]. Onboard sensing is range-limited and can be obstructed, while infrastructure connectivity is assumed available and is modelled at the application level. The Fairy has no privileged view of the environment: its only input is a CAM/CPMderived S-LDM state describing connected and sensed road users, which is what makes the recipient’s reported awareness the quantity the service can act on, rather than its actual perception. In the experiments, the simulator’s ground truth is indeed never provided. Objects are associated by vehicle identifier and stale entries are handled by the S-LDM. In our implementation, a local LDM cleaner runs every 0.5 s and removes entries older than 1 s; Fairy-supplied objects use the same local-LDM aging path after insertion. Fig. 1 summarizes the information flow and the separation between infrastructureside information selection and vehicle-local maneuver control. The ITS Fairy. Fig. 2 states the procedure executed by the ITS Fairy, and its line labels are cited below. For recipient v, let Bv (t) be fresh S-LDM vehicle objects within the implementation’s dynamic neighborhood, whose radius is max{∥vv ∥Th , dbase } with prediction horizon Th = 10 s and minimum neighborhood radius dbase = 5 m, following [27].
Fairy on updated recipient v F1: B ← fresh S-LDM neighbors of v F2: D ← objects in v’s CPM-derived history F3: for each o ∈ B \ D do F4: (t∗ , d∗ ) ← TCA(v, o) F5: if d∗ < dF = 5 m then F6: send ObjectState(o) to v Recipient on ObjectState(o) R1: map o to a local id and pose, insert into the local LDM R2: recompute local TCA and apply the unchanged maneuver policy Fig. 2. Selective missing-object assistance. F-steps run at the infrastructure, R-steps at the recipient. The Fairy decides what state to supplement; collisionrisk recomputation and control remain at the recipient.
Let Dv (t) be the S-LDM history of objects reported in v’s CPMs. The proof-of-concept assumes this reported set corresponds to the objects in v’s local LDM, yielding the missing set (F1–F3) Mv (t) = Bv (t) \ Dv (t).
(1)
For each o ∈ Mv (t), the Fairy applies a 2-D constantvelocity TCA relevance calculation (F4) with a 10 s prediction horizon. With relative position r and velocity let u, t∗ = clip(−r·u/(u·u), 0, Th ), with the zero-relative-velocity case handled separately; hence, the predicted separation at that instant is d∗ = ∥r + u t∗ ∥. An object is selected when its predicted separation at t∗ is below a fixed Fairy relevance threshold dF = 5 m (F5) (independent of dbase ), so that RiskTCA (v, o, t) = 1[d∗ < dF ], giving Av (t) = {o ∈ Mv (t) : RiskTCA (v, o, t) = 1}.
(2)
For each selected object (F6), the Fairy sends one custom CAM-like message carrying the recipient identifier, object station identifier, position/elevation, heading, speed, timestamps, dimensions, and station type; it sends no acceleration, confidence, or TCA verdict. The receiver derives local identifier and relative pose, inserts the state into its LDM, and recomputes TCA (R1–R2). The unchanged local controller also uses constant-velocity TCA over 10 s, but applies it to every object in its local LDM rather than to a speeddependent neighborhood. One message is sent per selected object, without acknowledgments/retransmissions, in line with how CAMs and CPMs are normally disseminated [2], [3]. III. E XPERIMENTAL M ETHODOLOGY Setup and Measurement. This section describes how we evaluate RQ1 and RQ2. The same comparison answers both questions: an unchanged vehicle-local collisionavoidance mechanism, operating on its local LDM, is run with and without Fairy assistance. We compare local-only and Fairy-assisted operation in the same SUMO–ms-van3t–SLDM co-simulation [1], [28]. ms-van3t [28], since extended as VaN3Twin [29], generates and encodes CAMs/CPMs, but the evaluated path carries them over UDP/IPv4 and virtual Ethernet to a UDP-to-AMQP relayer and then to the SLDM, which is subscribed to an AMQP broker [1]; the ITS Fairy uses the same broker to return the state. No vehicular wireless PHY/MAC or propagation model is instantiated, and
TABLE II E XPERIMENTAL MATRIX . F REQUENCIES DENOTE THE FAIRY ANALYSIS RATE , NOT THE CAM/CPM GENERATION FREQUENCY.
(a) Lane merge with blocked sensing LOS.
(b) Four-way intersection with four occluded approaches. Fig. 3. Evaluated sensing-occluded geometries. Buildings obstruct onboard sensing while infrastructure connectivity is retained.
vehicle applications deliberately do not consume direct peer V2V messages. The robustness parameter is therefore an independent application-level omission applied immediately before Fairy-state insertion into the local LDM, not a measured or simulated wireless Packet Delivery Ratio (PDR). This isolates the mechanism’s dependence on assistance availability from the propagation and congestion effects that cause loss in deployment; it does not approximate correlated wireless loss. The primary metric is the per-encounter minimum TCA, i.e., the smallest TCA value the vehicle-local detector produces during an encounter, considering the vehicles participating in that encounter. It is only defined once the detector holds both objects, so it measures the margin available after the conflict becomes locally computable. A vehiclelocal determineRisk() function runs every 100 ms; a separate AMQP thread subscribes to a Fairy-message topic, the simulator polls it every 10 ms, and successful insertion triggers immediate TCA reevaluation. This rate is fixed across all conditions and independent of the Fairy analysis rate swept in RQ2, where the Fairy queries the S-LDM for updated stations every 100/200/1000 ms for 10/5/1 Hz. Route traversal time from spawn to completion is a secondary metric. Experimental Design. We consider two main scenarios: lane merge and four-way intersection. The lane merge considers equal-length lanes separated by a building that blocks the 50 m vehicle onboard sensing LOS; the building affects perception, but not the communication path to the infrastructure. Vehicles reach the merging point approximately together; the closer one has priority, with a tie-break based on vehicle identifiers. The four-way scenario considers four simultaneously spawned straight-through vehicles at 30–80 km/h. Buildings obstruct sensing, vehicles yield to traffic on the right, and lower IDs break the four-way tie; route-dependent traversal times are reported separately. Fig. 3 shows both scenarios and road geometries. The nominal sweep is 30–120 km/h and 140– 200 km/h is an extreme-speed stress regime.
Experiment
Speeds (km/h) Assistance omission Frequency
Lane merge Stress regime Robustness Four-way
30–120 140–200 90/100/120 30–80
0% 0% 0–80% 0%
10 Hz 10 Hz 10/5/1 Hz 10 Hz
This design has known limitations that bound how the results should be read. The route generators schedule 100 sequential encounters per speed and configuration within one simulation execution: two vehicles per lane-merge encounter and four per intersection encounter. Routes and nominal speed are fixed, and our implementation sets a SUMO seed equal to 10, hence these are not independently seeded replications; therefore, we emphasize descriptive distributions. All configurations contain 100 encounters except the 10 Hz, 120 km/h cells at 40% and 70%, with 99 and 92 usable encounters. SUMO collision reports serve only as supporting evidence, because simulator hitboxes differ from the controller geometry. The robustness tests at 90, 100, and 120 km/h drop each Fairy-to-vehicle assistance message independently at the configured omission rate, leaving CAM/CPM reporting to the SLDM unaffected. With 10 Hz Fairy analysis, the omission rate takes the values 0, 5, 10, 15, 20, 30, 40, 50, 60, 70, and 80%; the 5 Hz and 1 Hz sweeps use the same values except 15%. Table II lists the full configuration set. IV. R ESULTS Effect of missing-object assistance. Fig. 4 shows the primary minimum-TCA results. In the lane merge, the localonly median is 0.46 s at 30 km/h and at most 0.08 s from 40–120 km/h. Fairy-assisted medians remain 7.47–9.47 s from 40–70 km/h, then decrease to 6.78, 5.76, 5.15, and 3.83 s at 80, 90, 100, and 120 km/h. The stress regime identifies the limit rather than a deployment range: assisted medians decline to 3.24, 2.56, 1.85, and 1.33 s at 140, 160, 180, and 200 km/h. Across every evaluated lane-merge speed, the smallest assisted minimum-TCA value exceeds the largest local-only value in the obtained results. Lane-merge traversal time is also lower over the nominal sweep: from 30 to 120 km/h, the mean reduction grows from about 11% to 45–47% on the two equallength routes. Our analysis attributes this to the avoidance style rather than to higher speeds being sustained: local-only vehicles detect the conflict late and brake hard, frequently to a full stop, whereas assisted vehicles adjust speed gradually. Because vehicles were configured to continue after stopping, local-only traversal times also include encounters that ended in a simulator collision. Above 140 km/h the traversal-time difference narrows; we report this as a coincident trend rather than attributing it to the TCA change. The four-way intersection shows the same effect in a multivehicle geometry. At 30 km/h, the local-only median minimum TCA is 1.25 s versus 4.83 s with assistance. From 40–80 km/h,
(a) Sensing-occluded lane merge.
(b) Sensing-occluded four-way intersection. Fig. 4. Per-encounter minimum TCA for local-only and Fairy-assisted operation (100 sequential encounters per speed and configuration in the main sweeps); the axis title abbreviates this as TCA. Local-only boxes are flattened near zero at most speeds. Speeds above 120 km/h in (a) are an extreme-speed stress regime.
local medians are only 0.01–0.03 s, which is deemed too low to effectively avoid collisions. Indeed, actuation phase delays on the order of a few tenths of a second have been recently measured in experimental automated vehicles [30]. Conversely, assisted medians decrease gradually from 4.65 to 2.24 s. Moreover, the lowest assisted minimum TCA among all 100 sequential encounters is still 4.39, 3.83, 2.73, 2.31, and 2.00 s at 40, 50, 60, 70, and 80 km/h, respectively; at every vehicle speed, the assisted value also exceeds the largest local-only value. Additionally, route-specific traversal times generally improve. Across the three routes other than Route 4, mean traversal-time reductions span 12.3–50.2% over 30– 80 km/h. We observed that Route 4 shows a smaller reduction and reverses at 80 km/h (19.78 s assisted versus 18.75 s local) ; because the precise cause of this reversal was not disentangled, we report it without attributing it to the Fairy. Robustness to missed assistance opportunities. Fig. 5 shows the results for the assistance-omission sweeps. At 10 Hz, the median minimum TCA changes modestly as the configured independent application-level assistance omission increases. Even at 80% omission, the median minimum TCA
is 5.49, 4.68, and 3.61 s at 90, 100, and 120 km/h respectively, versus local-only medians of 0.01, 0.04, and 0.08 s. At 5 Hz the corresponding medians remain similar, but the lower tail worsens; at 120 km/h the first quartile falls from 2.29 s to 1.36 s. At 1 Hz the omission sensitivity is much stronger: with 80% omission, the median remains 2.29 s at 90 km/h but falls to 0.04 and 0.08 s at 100 and 120 km/h, essentially matching local-only operation (Table III). The 1 Hz distributions are also visibly bimodal: individual encounters either retain most of the assisted margin or collapse onto the local-only value, so the median shifts abruptly rather than degrading smoothly and understates the spread between the two groups. For these three tested speeds, the median benefit therefore persists even at 80% configured omission with 10–5 Hz analysis, whereas the 1 Hz data show greater sensitivity to missed assistance opportunities. Quartiles characterize this bimodal regime only partially. It is worth noting how these configured independent omissions characterize application-level sensitivity, and not a tolerable packet loss rate for a real wireless link. Interpretation. RQ1. The comparison supports an information-availability reading rather than a better-predictor
(a) 10 Hz
(b) 5 Hz
(c) 1 Hz Fig. 5. Lane-merge per-encounter minimum-TCA distributions across the archived Fairy-to-vehicle assistance-omission sweeps at 10, 5, and 1 Hz analysis frequency. The local-only distributions are omission-independent and are drawn once, at the 0% position, as a reference. The 15% cell is absent at 5 and 1 Hz. At 1 Hz and high omission the median line coincides with the lower box edge; Table III gives those values numerically.
TABLE III M EDIAN MINIMUM TCA AT 80% FAIRY- TO - VEHICLE ASSISTANCE OMISSION ; FIRST QUARTILE IN PARENTHESES .
Fairy 10 Hz Fairy 5 Hz Fairy 1 Hz Local-only
90
100
120 km/h
5.49 (5.11) 5.44 (4.99) 2.29 (0.02) 0.01 (0.01)
4.68 (4.32) 4.52 (2.92) 0.04 (0.03) 0.04 (0.04)
3.61 (2.29) 3.51 (1.36) 0.08 (0.03) 0.08 (0.08)
reading. Recipient-side prediction, thresholds, maneuver logic and vehicle dynamics are unchanged; the assisted vehicle differs only in receiving selected S-LDM object state and recomputing TCA locally. The shape of the separation supports this. At every evaluated lane-merge speed the smallest assisted minimum TCA exceeds the largest local-only value, so the two distributions are disjoint rather than shifted: a difference in predictor quality would be expected to move a distribution, not to separate it. The assisted medians also decline smoothly with speed, from 7.47–9.47 s at 40–70 km/h to 3.83 s at 120 km/h, while local-only medians stay pinned at or below 0.08 s across 40–120 km/h. A margin that erodes gradually with closing speed while the unassisted baseline remains flat is the signature
of earlier information, not of a different decision rule. The metric must be read with its construction in mind. Minimum TCA comes from the same local detector that drives the controller, and under occlusion that detector cannot evaluate a conflict until the other vehicle enters the 50 m sensing model, which is why local-only values sit near zero. Part of the separation therefore reflects when the conflict became locally computable, the deficit the Fairy removes, so local-only values are not an independent measure of physical proximity, and a near-zero value does not distinguish a collision from a vehicle that stopped just in time. The minimum TCA is an operational conflict-imminence measure: the comparison gives evidence about information availability and the margin left to the controller, and it is not a collision-rate result. RQ2. The frequency–omission interaction follows from how much road is covered between successive analyses. At 10 Hz the Fairy re-examines a recipient every 100 ms; at 1 Hz once per second, which at 120 km/h corresponds to roughly 33 m of distance travelled against a 50 m sensing model. The observed behavior matches the obtained results. At 10 Hz the benefit survives 80% omission almost intact, with medians of 5.49, 4.68 and 3.61 s at 90, 100 and 120 km/h against local-only medians of 0.01, 0.04 and 0.08 s; even at 80% omission a
recipient still receives about two assistance opportunities per second on average, so one typically arrives before the conflict becomes critical. At 1 Hz the same omission leaves 2.29 s at 90 km/h but 0.04 and 0.08 s at 100 and 120 km/h, and the collapse is abrupt rather than gradual. The 1 Hz distributions are correspondingly bimodal: individual encounters either retain most of the assisted margin or fall back to the local-only value. This bimodality is what a delivery-opportunity mechanism predicts: an object either reaches the recipient before the conflict becomes critical or contributes almost nothing. V. C ONCLUSION The ITS Fairy turns a broader centralized view, based on a Server Local Dynamic Map (S-LDM), into recipient-specific assistance by sending only conflict-relevant object state absent from a vehicle’s reported awareness, leaving collisionrisk recomputation and maneuver control local. Reading that awareness from the recipient’s own CPMs, rather than inferring it from geometry or predicted perceptibility, costs no additional message, besides the ones an aggregation point is already receiving. In sensing-occluded lane merging, assisted minimum TCA holds at 3.83–9.47 s where the same controller without assistance stays at or below 0.08 s from 40–120 km/h, and mean traversal time falls by about 11% to 45–47% across that range. The benefit survives 80% assistance omission at 10 Hz analysis but not at 1 Hz, indicating that what matters is delivery opportunity rather than delivery volume. Standardized messaging, measured delay and age of information, and controlled-seed replication are the natural next steps. R EFERENCES [1] C. M. Risma Carletti, F. Raviglione, C. Casetti, F. Stoffella, G. M. Yilma, and F. Visintainer, “S-LDM: Server local dynamic map for 5G-based centralized enhanced collective perception,” Vehicular Communications, vol. 49, p. 100819, 2024. [2] European Telecommunications Standards Institute (ETSI), “Intelligent transport systems (ITS); vehicular communications; basic set of applications; part 2: Specification of cooperative awareness basic service,” ETSI, Tech. Rep. EN 302 637-2, Nov. 2014. [3] ——, “Intelligent transport system (ITS); vehicular communications; basic set of applications; collective perception service; release 2,” ETSI, Tech. Rep. TS 103 324, Jun. 2023. [4] F. Hawlader, F. Robinet, and R. Frank, “Cooperative perception using V2X communications: An experimental study,” in IEEE VTC2024-Fall, 2024, pp. 1–7. [5] H. Pinho and J. Ferreira, “Sensor fusion for improved cooperative perception in CCAM,” in IEEE VTC2024-Spring, 2024, pp. 1–5. [6] J. Götz, L. Mathuseck, L. Busch, and K. David, “On the influence of the communication delay in context-aware collision avoidance systems,” in IEEE VTC2024-Spring, 2024, pp. 1–5. [7] Q. Huang, J. Zeng, C. Guo, J. Liu, and H. Pan, “AoI analysis for automatic repeat-request in vehicular cooperative perception networks,” in IEEE VTC2024-Fall, 2024, pp. 1–5. [8] T. Karunathilake and A. Förster, “Cutting the clutter: Optimizing CPM processing in VANETs on the receiver side,” in IEEE VTC2025-Fall, 2025, pp. 1–7. [9] M. Khoshkdahan, J. P. Arockiasamy, A. F. Comeca, and A. Vinel, “Cooperative robotics reinforced by collective perception for traffic moderation,” in IEEE VTC2026-Spring, 2026. [10] Y. Miyata and H. Shigeno, “Collective perception service with risk and redundancy-based object selection in cellular-V2X,” in ICMU, 2025, pp. 1–6.
[11] P. Zhou, P. Kortoçi, Y. Yau, B. Finley, X. Wang, T. Braud, L. Lee, S. Tarkoma, J. Kangasharju, and P. Hui, “AICP: augmented informative cooperative perception,” IEEE Trans. Intell. Transp. Syst., vol. 23, no. 11, pp. 22 505–22 518, 2022. [12] C. Ma, H. Li, K. Long, H. Zhou, Z. Liang, P. Li, H. Yu, and X. Li, “Realtime identification of cooperative perception necessity in road traffic scenarios,” Transportation Research Part C: Emerging Technologies, vol. 185, p. 105547, 2026. [13] J. Liu, C. Ma, H. Zhou, W. Tang, S. Liang, H. Ding, X. Li, and B. Ran, “SRA-CP: Spontaneous risk-aware selective cooperative perception,” arXiv preprint arXiv:2511.17461, 2025. [14] J. Liu, J. Wu, X. Cui, T. Huang, D. Zhang, and Y. Bu, “Risk-guided scheduling for spatio-temporal collaborative perception in vehicular networks,” in IEEE INFOCOM 2026. [15] S. Manjunatha, S. S. Avedisov, O. Altintas, and A. H. Sakr, “Experimental insights into intent sharing for enhanced infrastructure-assisted cooperative perception,” IEEE Intell. Transp. Syst. Mag., vol. 18, no. 5, pp. 74–84, 2026. [16] H. Qiu, P.-H. Huang, N. Asavisanu, X. Liu, K. Psounis, and R. Govindan, “AutoCast: Scalable infrastructure-less cooperative perception for distributed collaborative driving,” in MobiSys, 2022, pp. 128–141. [17] B. Hakim, S. Sorour, M. S. Hefeida, W. S. Alasmary, and K. H. Almotairi, “CCPAV: Centralized cooperative perception for autonomous vehicles using CV2X,” Ad Hoc Networks, vol. 142, p. 103101, 2023. [18] P. Lv, J. Han, Y. He, J. Xu, and T. Li, “Object perceptibility prediction for transmission load reduction in vehicle-infrastructure cooperative perception,” Sensors, vol. 22, no. 11, p. 4138, 2022. [19] T. Higuchi, M. Giordani, A. Zanella, M. Zorzi, and O. Altintas, “Valueanticipating V2V communications for cooperative perception,” in IV 2019, pp. 1947–1952. [20] M. K. Abdel-Aziz, C. Perfecto, S. Samarakoon, M. Bennis, and W. Saad, “Vehicular cooperative perception through action branching and federated reinforcement learning,” IEEE Trans. Commun., vol. 70, no. 2, pp. 891–903, 2022. [21] Y. Hu, S. Fang, Z. Lei, Y. Zhong, and S. Chen, “Where2comm: Communication-efficient collaborative perception via spatial confidence maps,” in NeurIPS, vol. 35, 2022. [22] J. Cui, H. Qiu, D. Chen, P. Stone, and Y. Zhu, “Coopernaut: Endto-end driving with cooperative perception for networked vehicles,” in IEEE/CVF CVPR, 2022, pp. 17 231–17 241. [23] L. Westhofen, C. Neurohr, T. Koopmann, M. Butz, B. Schütt, F. Utesch, B. Neurohr, C. Gutenkunst, and E. Böde, “Criticality metrics for automated driving: A review and suitability analysis of the state of the art,” Archives of Computational Methods in Engineering, vol. 30, pp. 1–35, 2023. [24] Y. Asabe, E. Javanmardi, J. Nakazato, M. Tsukada, and H. Esaki, “Enhancing reliability in infrastructure-based collective perception: A dual-channel hybrid delivery approach with real-time monitoring,” IEEE Open J. Veh. Technol., vol. 5, pp. 1124–1138, 2024. [25] J. Thunberg, D. Bischoff, F. A. Schiegg, T. Meuser, and A. Vinel, “Unreliable V2X communication in cooperative driving: Safety times for emergency braking,” IEEE Access, vol. 9, pp. 148 024–148 036, 2021. [26] S. Ren, Z. Lei, Z. Wang, M. Dianati, Y. Wang, S. Chen, and W. Zhang, “Interruption-aware cooperative perception for V2X communicationaided autonomous driving,” IEEE Trans. Intell. Veh., vol. 9, no. 4, pp. 4698–4714, 2024. [27] M. Malinverno, G. Avino, C. Casetti, C. F. Chiasserini, F. Malandrino, and S. Scarpina, “Performance analysis of c-v2i-based automotive collision avoidance,” in 19th IEEE International Symposium on ”A World of Wireless, Mobile and Multimedia Networks”, WoWMoM 2018, Chania, Greece, June 12-15, 2018. IEEE Computer Society, 2018, pp. 1–9. [28] F. Raviglione, C. M. Risma Carletti, M. Malinverno, C. Casetti, and C. F. Chiasserini, “ms-van3t: An integrated multi-stack framework for virtual validation of V2X communication and services,” Computer Communications, vol. 217, pp. 70–86, 2024. [29] R. Pegurri, D. Gasco, F. Linsalata, M. Rapelli, E. Moro, F. Raviglione, and C. Casetti, “Van3twin: The multi-technology V2X digital twin with ray tracing in the loop,” IEEE Trans. Wirel. Commun., vol. 25, pp. 16 812–16 827, 2026. [30] F. Werner, T. Heintzenberg, M. Lienkamp, and J. Betz, “Drifting in the future: Stabilizing path following drifting on high-latency vehicle systems,” CoRR, vol. abs/2606.27914, 2026.