TruckV2X: Cooperative Perception for Heavy Trucks

• TruckV2X is a joint communications-perception infrastructure tailored for heavy-duty trucks, addressing unique challenges like occlusion and blind spots.
• It leverages multiple V2X standards and advanced localization methods to support critical use cases such as platooning, lane merge assistance, and convoy safety.
• The TruckV2X dataset and benchmarks demonstrate strong cooperative perception performance, revealing asymmetric occlusion recovery rates among tractor, trailer, CAV, and RSU.

Searching arXiv for recent and directly relevant papers on TruckV2X, V2X standards, and cooperative truck perception. TruckV2X denotes vehicle-to-everything connectivity, localization, and perception tailored to the operational realities of heavy-duty trucks and freight systems, and in recent literature it also names a truck-centered cooperative perception dataset built around tractors, trailers, connected autonomous vehicles, and roadside units (Dettinger et al., 2024, Xie et al., 13 Jul 2025). As a systems concept, TruckV2X spans V2V, V2I, V2N, and V2P communication; integrated sensing and communication for full-pose truck estimation; corridor and yard automation; and multi-agent cooperative perception intended to overcome the extensive blind spots and occlusions created by large articulated vehicles (Führling et al., 2024). Heavy trucks intensify the central V2X design constraints—latency, reliability, range, occlusion handling, interoperability, and scalability—because their large dimensions increase self-occlusion and shadowing, convoys create dense clusters, and longer stopping distances require more conservative communication and perception margins than in light-vehicle settings (Dettinger et al., 2024, Shah et al., 2023).

1. Scope and operational definition

TruckV2X is grounded in the general V2X model in which vehicles connect with other vehicles, infrastructure, pedestrians, and the network or cloud to extend perception and coordinate maneuvers, but it specializes that model for heavy trucks, freight corridors, depots, and platoons (Dettinger et al., 2024). The truck-relevant use cases identified in the literature include platooning and convoy operations, cooperative adaptive cruise control, lane merge assistance, blind-spot and angle detection, collision and queue avoidance, freight-corridor operations with roadside units, and yard or depot automation. The same literature also treats trucks as particularly demanding V2X nodes because long tractors and trailers create extensive lateral and rear blind zones, trailers pivot dynamically during turns, and large metal bodies alter both communication geometry and perception coverage (Xie et al., 13 Jul 2025).

A recurrent mistake is to regard TruckV2X as generic passenger-car V2X with larger antennas or higher transmit power. The technical literature instead models the truck itself as a distinct cyber-physical platform whose geometry, articulation, and operating regime materially change observability, communications, and safety requirements. This is evident in rigid-body localization models that estimate truck translation and rotation rather than a single reference point, in cooperative perception pipelines that exploit infrastructure vantage points to compensate for trailer-induced occlusions, and in congestion-control strategies that prioritize long-range reliability and stable inter-reception times under dense convoy traffic (Führling et al., 2024, Tan et al., 2023).

TruckV2X is therefore best understood as a joint communications-perception infrastructure for heavy vehicles. Vehicle-side sensing and computation, RSU-side measurement and fusion, and network or cloud orchestration are treated as co-designed components rather than as separable subsystems. A plausible implication is that “truck connectivity” and “truck perception” are no longer cleanly separable categories in current research, because the same signals increasingly support both data exchange and sensing.

2. Communication stacks, KPIs, and deployment architecture

The principal communications families discussed for TruckV2X are IEEE 802.11p-based DSRC or ITS-G5 and the cellular family comprising LTE-V2X and NR-V2X. DSRC and ITS-G5 use IEEE 802.11p PHY/MAC with OCB mode and CSMA/CA; IEEE 802.11bd adds MIMO, LDPC, 256-QAM, BPSK-DCM, retransmissions, and optional 60 GHz extensions and localization. LTE-V2X Rel-14/15 introduces sidelink PC5 for direct V2V, V2I, and V2P, together with Uu for V2N; NR-V2X Rel-16/17 adds unicast and groupcast alongside broadcast, sidelink feedback, flexible numerology, short TTIs, and scheduled and autonomous sidelink modes. LTE-V2X and NR-V2X are not air-interface compatible, so coexistence relies on in-device TDM/FDM and cross-RAT control (Dettinger et al., 2024).

For heavy trucks, these differences are operationally consequential. Highway platooning and dense freight corridors benefit from centralized scheduling in LTE-V2X Mode 3 or NR-V2X SL Mode 1, while advanced driving and lane merge assistance benefit from NR-V2X groupcast or unicast and feedback. Cooperative perception and blind-spot mitigation impose higher rate requirements than Day-1 safety messaging, and the survey consolidates corresponding targets: platooning requires latency of 10–500 ms depending on speed and convoy size, reliability of 90–99.99%, and V2V data rate of approximately 2–3 Mbps; advanced driving requires latency of 3–100 ms, reliability of 90–99.99%, and V2V/V2I rates of about 0.5–10 Mbps; extended sensors and cooperative perception require latency of 3–50 ms, reliability of 95–99.99%, V2N rates of 10–50 Mbps for map sharing, and V2V rates of 25–1000 Mbps for collective perception bursts. End-to-end latency is commonly decomposed as

$$T_{e2e} = T_{proc} + T_{access} + T_{tx} + T_{prop} + T_{sched},$$

with the access and scheduling terms depending on whether the system uses CSMA/CA, SP-SPS, or centralized scheduling (Dettinger et al., 2024).

TruckV2X architectures are generally hybrid. RSUs along freight corridors broadcast MAP, SPaT, and TIM and can aggregate collective perception; PC5 sidelink preserves convoy cohesion independent of coverage; and Uu or V2N provides backend coordination and over-the-air updates. MEC is repeatedly positioned as a suitable placement for cooperative perception fusion, convoy management, and yard automation when latency budgets permit. In network-assisted freeway V2I, a representative LTE-A study used RSUs offset 35 m from a six-lane highway and spaced 1732 m apart; at 10 MHz and 128 kb/s per vehicle, empirical performance suggested that a truck-dominated cell should target roughly 40–60 trucks per RSU to keep about 95–99% of trucks at at least 128 kb/s downlink under LMMSE plus precoding, while higher periodicity or payloads require either more bandwidth or denser RSU placement (Luoto et al., 2016).

A practical conclusion in the literature is that no single stack dominates all truck use cases. ITS-G5 and 802.11bd remain attractive for very low latency at short range, whereas LTE-V2X and especially NR-V2X are preferred for reliable platooning, convoy coordination, and cooperative perception under high density (Dettinger et al., 2024).

3. Localization, sensing, and cooperative perception

A defining technical strand in TruckV2X is the shift from point localization to full-pose estimation of a heavy vehicle as a rigid body. Wireless rigid body localization extends wireless localization from single-point targets to three-dimensional objects modeled as rigid bodies: instead of estimating one position, it estimates translation and rotation of the whole object and leverages fixed shape constraints. In this model, a rigid body with body-frame points $\mathbf{c}_k$ is mapped to transformed points $\mathbf{s}_k$ through a rotation $\mathbf{R}$ and a translation $\mathbf{t}$, while the fixed inter-point Euclidean distances in the conformation matrix $\mathbf{C}$ remain invariant. Communication-derived measurements such as range, angle of arrival, angle difference of arrival, time-of-arrival, and RSSI are then fused with the known body shape. For frame-to-frame tracking, the instantaneous velocity of a sensor point is modeled as

$$\dot{\boldsymbol{s}_i}=[\boldsymbol{\omega}]^\times\boldsymbol{R}\boldsymbol{c}_i+\dot{\boldsymbol{t}}_i,$$

which directly supports time-evolving convoy and platooning scenarios (Führling et al., 2024).

The heavy-truck case gives this formulation unusual practical importance. Large bodies and trailers block line of sight, so incomplete observations are common; the rigid-body constraint permits reconstruction of missing cross-distances through matrix completion over a hollow squared Euclidean distance matrix. The same work reports that, in simulated 3D rigid-body localization via least squares, two sensors are insufficient, four sensors are sufficient to locate a 3D rigid body with large performance gain, and errors decrease and begin to converge as the number of sensors grows to ten. Angle measurements are emphasized as particularly informative for rigid bodies because different body points may share identical ranges to an anchor while still differing in angle, which sharpens pose observability when the shape is known. For trucks, this directly supports docking, lane merging, tight-yard maneuvering, convoy tracking, blind-spot mitigation, and relative localization in low-infrastructure settings (Führling et al., 2024).

TruckV2X perception also relies on infrastructure-centered multi-agent fusion. In Adaptive Road-to-Vehicle Perception, an RSU supplies a stable, wide-range vantage point and acts as a communication hub, while vehicles share LiDAR-based BEV features and poses. The method combines a Dynamic Perception Representing module, which constructs a directed collaborative graph over agents, with a Road-to-Vehicle Perception Compensating module that residually injects RSU information when vehicle-side features under-represent dynamic or occluded actors. On V2X-Sim, AR2VP reports segmentation mIoU of 85.05%, [email protected] of 94.50%, and [email protected] of 92.77%, while its RSU experience-replay variant reports detection mAP@0.5 / Forget of 89.52 / 6.16 and segmentation mIoU / Forget of 71.51 / 11.48 under inter-scene changes. Under 32× feature compression, segmentation mIoU drops by only 1.31% yet remains above the Disco baseline, which positions RSU-centered fusion as a bandwidth-efficient response to dynamic truck scenes (Tan et al., 2023).

At mmWave and sub-THz frequencies, TruckV2X must also manage blockage explicitly. Highway blockage analysis models vehicles as 3D boxes with length, width, and height, and defines blockage through a Fresnel-based height criterion. Taller, wider, and longer trucks increase blockage probability, while rooftop antenna placement improves service probability; the reported results show an SNR difference of approximately 4.2 dB favoring rooftop over bumper mounting. IOS-enabled sensing-assisted communication extends the same logic at 30 GHz by mounting an intelligent omni-surface on the vehicle so that one part of the incident RSU signal is reflected to strengthen echoes for sensing and another part is refracted into the cabin for communication. In a representative trajectory, the achievable rate is maximized at about $\eta^\ast \approx 0.18$ and $\beta^{R\ast}\approx 0.8$, and the scheme shows pronounced throughput advantage over prediction-only or random-reflection baselines at low transmit power (Dong et al., 2021, Meng et al., 2022).

4. Congestion control, scalability, and reliability engineering

TruckV2X pushes shared-spectrum systems toward congestion limits because periodic convoy traffic coexists with broader awareness messaging and, increasingly, collective perception. In an HDV platooning study at 5.9 GHz ITS-G5, CAMs of 285 bytes and unicast Platooning Control Messages of 301 bytes at 2 Hz were simulated on a four-lane-per-direction highway. The paper reports that moderate-to-high-density scenarios drive the channel busy ratio toward convergence near 0.65, with a theoretical limit around 0.68, and that hidden and exposed nodes aggravate collisions. Offloading intra-platoon PCM traffic to bumper-to-bumper radar-based communication substantially reduces shared-channel stress: at 0% RadCom penetration, PDR was 0.6985, S-CBR 0.6176, and latency 136.80 ms; at 50%, PDR was 0.7859, S-CBR 0.6119, and latency 109.57 ms; at 100%, PDR was 0.9015, S-CBR 0.2217, and latency 1.45 ms. For a four-truck platoon, the reported fuel reductions were 2% at 50% penetration and 5.6% at 100%, driven by shorter safe gaps (Haller et al., 2024).

On cellular sidelink, the principal scalability pathology is persistent collision under sensing-based semi-persistent scheduling. In LTE-V2X Rel-14 PC5 Mode-4, a vehicle senses a 1 s window, predicts the next selection window, and reserves resources semi-persistently without re-evaluation; under heavy density this creates repeated collisions and long information-age tails. The corresponding study reports that, with 10 MHz and no congestion control, PRR remains above 90% within 100 m even in heavy density but can fall to approximately 60% at 200 m. Three remedies are emphasized. First, 20 MHz channelization materially lifts PRR across distances and densities. Second, rate-centric congestion control in SAE J3161/1 improves PRR by up to 50% for 0–200 m pairs in heavy density, while reducing information age by at least 500 ms for 99.9% of pairs at 0–200 m and by at least 1300 ms for 99% of pairs at 200–300 m. Third, one-shot transmissions inserted every 2–6 periods break persistent collision streaks; at 200–300 m in heavy density, one-shot reduces information age by more than 1500 ms at the 99.9% point compared with rate-only or rate-plus-power control (Shah et al., 2023).

Reliability can also be raised by radio diversity. A multi-RAT study duplicates packets concurrently over LTE-Uu and PC5 so that a receiver accepts the earliest successfully received copy. In dense urban conditions at 200 m range and 1000 vehicles/km², PC5 plus LTE-Uu multicast raises PRR from about 90% for PC5 alone to about 97% with duplication; at 300 m range and 500 vehicles/km², where both single-RAT PRRs are below 85%, the same duplication pushes PRR above 95%. The analytical interpretation is straightforward: if the two paths are sufficiently decorrelated, combined reliability follows

$$R_{\mathrm{combined}} = 1 - (1 - R_{\mathrm{Uu}})(1 - R_{\mathrm{PC5}}),$$

while latency is governed by the minimum of the two arrival processes (Lianghai et al., 2018).

Network slicing offers another separation mechanism. In a highway LTE-A evaluation, an autonomous driving slice for safety messaging and an infotainment slice for video streaming were isolated by placing them on different carriers, with the autonomous driving slice served by vehicle-based virtual RSUs chosen through spectral clustering. In dense traffic with inter-vehicular distance 1–100 m, safety PRR improved from 31.15% for the baseline RSU architecture to 99.47% with network slicing at $\mathbf{c}_k$0 m; corresponding results for 100–200 m and 200–300 m were 99.65% and 99.78%, respectively. The same study showed a trade-off when infotainment relaying was added: infotainment throughput improved, but safety PRR fell relative to slicing alone because of added V2V interference (Khan et al., 2018).

Across these results, the engineering pattern is consistent. TruckV2X reliability under load is obtained not by a single MAC improvement but by structural separation of traffic classes, bandwidth enlargement, collision de-correlation, and, where needed, offloading of strictly local convoy traffic away from the shared safety channel.

5. Infrastructure services, security, and standardization

TruckV2X is not limited to convoy control and cooperative perception; it also encompasses corridor management, emergency interactions, and safety supervision through RSUs and base stations. A DSRC/WAVE emergency alert system field-tested at the TiHAN testbed equips emergency and commercial vehicles with OBUs based on a quad-core NXP i.MX6 platform, dual IEEE 802.11p radios, CAN, and BLE, while RSUs on traffic signals or lamp-posts relay alerts to a centralized Grafana-monitored base station. The system broadcasts SAE J2735 Basic Safety Messages every 100 ms, uses a 30 m threshold to trigger “provide pathway” alerts to vehicles ahead of an emergency vehicle, and reports operational range “upto 600m line of sight and drops to 350m without line of sight.” It also implements accident or emergency braking detection from the last seven speed packets, VRU detection through IP cameras and YOLOv5, and an intersection collision model based on projected collision point coordinates $\mathbf{c}_k$1, which is especially relevant for long trucks in blind turns (Kumar et al., 2024).

Security in TruckV2X is typically treated as layered rather than purely cryptographic. DSRC/WAVE uses IEEE 1609.2, while ETSI ITS security uses TS 102 097/940/941/942 with certificate-based authentication, pseudonym rotation, and misbehavior reporting. The literature also proposes intelligent V2X security as a physical-layer complement that adapts protection according to location, utility, application, environment, situation or time, and vehicle specifications. In that framework, high-security eavesdropping protection uses artificial noise, interference-based methods, or hybrid precoding; high-security spoofing protection uses joint CSI, RSS, and analog front-end features; and high-security anti-jamming combines multi-antenna methods with spread spectrum and higher processing gain. For trucks, the framework explicitly elevates security in platooning and hazardous-cargo cases and treats leader–follower control as a high-utility application (Furqan et al., 2019).

Standardization work places TruckV2X within a broader transition from B5G to 6G. Release-18 and beyond in 5G-Advanced include network-based positioning, sidelink-based positioning, hybrid positioning across UWB, GNSS, Bluetooth, and WiFi, and AI/ML enhancements under study. In the 6G IMT-2030 framing, sensing and advanced positioning are treated as key ISAC drivers, and rigid body localization is described as an interesting and required enhancement for ITS and automotive use cases. Regional deployment remains fragmented: in the United States, the FCC reallocated 5.9 GHz in 2020 and momentum has shifted toward C-V2X; in Europe, ITS-G5 and C-V2X share 5.875–5.925 GHz under coexistence mitigation, with mass deployment slipping from 2025 toward approximately 2027; and in China, policy and deployment focus on C-V2X and NR-V2X with strong 5G coverage (Führling et al., 2024, Dettinger et al., 2024).

A common misconception is that RSUs alone solve truck safety. The literature instead assigns complementary roles: trucks remain sensing and decision nodes, RSUs act as external sensors and communication hubs, and network or cloud layers coordinate or fuse information when timing budgets allow. Another misconception is that security can be bolted on afterward; in practice, certificate management, pseudonym rotation, and physical-layer adaptation all feed back into latency and processing budgets, including the $\mathbf{c}_k$2 term in the end-to-end delay decomposition (Dettinger et al., 2024).

6. The TruckV2X dataset, benchmarking, and research directions

In 2025, “TruckV2X” was introduced as the first large-scale truck-centered cooperative perception dataset designed specifically around articulated heavy-duty trucks and their occlusion patterns (Xie et al., 13 Jul 2025). The dataset models four cooperative agents—tractor, trailer, CAV, and RSU—because the tractor and trailer are treated as separate cooperative platforms rather than as a single ego vehicle. This design follows directly from truck kinematics: tractor–trailer combinations create large and persistent blind zones, and the authors’ analysis reports that during 90° turns more than 70% of a 30 m observational disk can be obscured. The dataset therefore positions the truck not only as an occlusion source but also as a “mobile perception platform” whose distributed sensing can help other agents recover occluded objects (Xie et al., 13 Jul 2025).

The sensing setup is multi-modal and synchronized at 10 Hz. Tractor and trailer each carry five RGB cameras and two 64-channel LiDARs; the CAV carries four RGB cameras and one 64-channel LiDAR; and the RSU carries one RGB camera and one 64-channel LiDAR, commonly mounted at about 4 m height. The paper reports 64 scenarios, 88,396 synchronized LiDAR-image frames totaling 149.6 GB, and 1.18 million annotated 3D boxes across seven classes. Category counts are given explicitly: Car 672,351; Pedestrian 229,041; Van 95,275; Truck 86,702; Bus 42,642; Bicycle 1,988; Motorcycle 1,968. The published train/val/test split is 38/9/17 scenarios with 13,243/2,839/6,017 frames, and the paper notes no explanation for the discrepancy between these split counts and the larger total, which likely reflects a difference between synchronized scene frames and multi-agent frames (Xie et al., 13 Jul 2025).

Its central analytical construct is the occlusion recovery rate, defined qualitatively as the fraction of one agent’s occluded objects that are visible to a cooperating agent. The results show asymmetry across partners. For the CAV, the truck provides strong recovery, with ORR above 0.8 in more than 60% of cases, while the RSU typically provides ORR around 0.4–0.6. For the truck, RSU assistance is slightly stronger than CAV assistance, at roughly 0.4–0.6 versus 0.2–0.4. For the RSU, the truck often yields ORR above 0.8, whereas CAV-to-RSU recovery is usually below 0.4. These numbers complicate a simple “infrastructure always dominates” narrative: elevated RSUs are important, but the truck itself often provides the strongest complementary viewpoint (Xie et al., 13 Jul 2025).

Benchmarking is centered on 3D detection over the BEV region $\mathbf{c}_k$3 m and $\mathbf{c}_k$4 m, using mAP at IoU thresholds 0.3, 0.5, and 0.7 for heavy vehicles, light vehicles, and VRUs, with inter-agent communication limited to 300 m. To reduce false positives in highly occluded frames, the benchmark removes boxes containing fewer than 5 LiDAR points for VRUs or 15 points for vehicles before evaluation. The paper states that eight benchmark implementations were provided, though it does not list them numerically in the supplied text (Xie et al., 13 Jul 2025).

The dataset also clarifies several open research directions. First, articulation-aware cooperative perception remains underdeveloped: current methods must explicitly handle turning thresholds such as $\mathbf{c}_k$5, where trailer-induced occlusions rise sharply. Second, communication constraints remain abstract in many perception benchmarks; the TruckV2X benchmark uses a 300 m communication cap but does not model latency, bandwidth, or packet loss, so perception algorithms still need systematic evaluation under realistic V2X network impairments. Third, the literature increasingly points toward joint localization-perception stacks in which rigid-body pose estimation, RSU-centric perception compensation, and convoy communications are trained or optimized together rather than independently. This suggests that future TruckV2X systems will likely treat tractor, trailer, RSU, and cooperating vehicles as a single distributed sensing graph rather than as loosely coupled autonomous nodes (Führling et al., 2024, Xie et al., 13 Jul 2025).