CoAnchor: Robust Collaborative Perception under Spatio-Temporal Misalignment via Object-Level Anchors

Chi Li; Rui Lin; Aobo Ji; Dongzhu Xu

cs.CVAug 21, 2026

CoAnchor: Robust Collaborative Perception under Spatio-Temporal Misalignment via Object-Level Anchors

Authors: Chi Li, Rui Lin, Aobo Ji, Dongzhu Xu

Abstract

Collaborative perception extends the sensing range of a single vehicle by fusing observations from nearby agents, which improves the robustness of autonomous driving. In realistic deployments, however, the received collaborator messages are often affected by both communication delay and relative-pose noise, which jointly cause stale observations, spatial misalignment, and unstable feature fusion. Existing methods usually address these issues from either the spatial or temporal side, but handling them jointly in a unified and efficient manner remains challenging. In this paper, we propose CoAnchor, an anchor-centric spatio-temporal alignment framework for asynchronous collaborative perception. Instead of directly reasoning on dense BEV features, CoAnchor builds sparse object-level spatio-temporal anchors as a shared interface for pose correction and tightly connects spatial refinement, temporal propagation, and current-time verification within one unified loop, while keeping the overall correction process lightweight. Extensive experiments on both simulated and real-world datasets illustrate that CoAnchor remains competitive under clean settings and improves the robustness under joint delay and pose perturbations with a favorable practical accuracy-efficiency trade-off.

Figures & tables

Figure 1. Temporal and spatial misalignment between ego vehicle and cooperative agents. Two panels illustrate cross-agent misalignment. The temporal panel shows ego and neighbor LiDAR observations captured at timestamps separated by a delay. The spatial panel shows a relative-pose error causing the two agents' point clouds and bounding boxes for the same pedestrian to be offset.

Figure 4. Proposed architecture of CoAnchor . Given delayed collaborator features, object cues, and noisy relative poses, CoAnchor constructs object-level anchors to refine the relative pose and to provide ego-guided priors for delay compensation. The propagated and verified anchor states are then fed back to later pose refinement and used to support subsequent collaborative fusion. Pipeline of CoAnchor. Single-agent perception extracts BEV features, object boxes, and poses; multi-agent communication supplies delayed neighbor features and object states. CoAnchor performs spatio-temporal anchor initialization, object temporal propagation, closed-loop posterior scoring with pose refinement, and anchor-guided feature fusion before the detection head.

Method	OPV2V ( Xu et al., 2022b )				V2V4Real ( Xu et al., 2023b )
Method	No Noise	P-Hard	J-Hard	D-Hard	No Noise	P-Hard	J-Hard	D-Hard
CoAlign ( Lu et al., 2023 )	96.64 / 91.31	86.49 / 74.19	85.90 / 72.82	83.19 / 71.37	72.64 / 42.97	51.34 / 32.81	53.59 / 33.22	58.26 / 34.26
ERMVP ( Zhang et al., 2024 )	96.59 / 94.01	63.61 / 48.57	57.46 / 42.95	47.86 / 36.56	78.38 / 52.55	51.63 / 34.69	53.16 / 34.55	59.81 / 36.81
V2X-ViT ( Xu et al., 2022a )	95.71 / 87.54	77.34 / 60.06	75.90 / 55.94	68.55 / 52.28	78.33 / 51.45	51.87 / 35.37	53.83 / 36.24	59.84 / 37.62
CoST ( Tang et al., 2025 )	97.20 / 93.70	51.00 / 32.10	47.90 / 27.80	42.60 / 26.90	74.10 / 50.00	49.50 / 30.60	51.80 / 32.00	58.00 / 35.80
TraF-Align (ego-only)	84.49 / 78.57	84.49 / 78.57	84.49 / 78.57	84.49 / 78.57	62.19 / 40.55	62.19 / 40.55	62.19 / 40.55	62.19 / 40.55

Table 1. Comparison with state-of-the-art methods on OPV2V and V2V4Real. Each entry is reported as [email protected] / [email protected].

Panel of Figure 5 — Figure 5. Robustness analysis under progressively increasing temporal and spatial perturbations. Panels (a)–(b) fix the delay at 100 ms on OPV2V and vary the pose-noise magnitude. Panels (c)–(d) fix the pose-noise standard deviations at $(0.3\,\mathrm{m},0.3^{\circ})$ on V2V4Real and vary the communication delay. [email protected] and [email protected] are reported separately. Four line charts report detection average precision as perturbations increase. The top row shows OPV2V AP at IoU thresholds 0.5 and 0.7 as pose noise grows under a fixed 100 millisecond delay. The bottom row shows V2V4Real AP at the same thresholds as delay grows under fixed pose noise. CoAnchor remains the highest and most stable curve, while most baselines decline more sharply.

Panel of Figure 6 — Figure 6. Qualitative analysis under the Joint-Hard setting. Four LiDAR bird's-eye-view detection maps compare CoAlign, V2X-ViT, baseST, and CoAnchor under the Joint-Hard setting. Yellow callouts enlarge crowded regions. CoAnchor shows closer overlap between the red and green boxes and fewer visibly shifted or duplicated detections than the comparison methods.

Variant	OPV2V
Variant	[email protected]	[email protected]
w/o historical initialization cues	92.47	81.59
w/o ego-guided correction	94.03	85.27
w/o anchor feedback	94.31	86.73
CoAnchor (Full)	95.15	87.96

Table 2. Module-wise ablation study on OPV2V.

Method	V2V4Real		Time (ms/frame)
Method	[email protected]	[email protected]	Time (ms/frame)
V2X-ViT ( Xu et al., 2022a )	53.83	36.24	110.62
TraF-Align ( Song et al., 2025 )	58.18	35.09	121.73
CoAnchor w/o feedback	66.19	43.35	49.39
CoAnchor +1 round feedback	67.04	44.05	51.62
CoAnchor +2 round feedback	66.96	44.16	52.51

Table 3. Performance and runtime analysis on V2V4Real.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Epochs	LR / scheduler	Perturbation during training
CoAnchor (single-agent)	30	Multi-step decay; initial lr $=0.002$ $\gamma=0.1$ , milestones [10, 20]	perfect setting throughout
CoAnchor CoAlign	30	Multi-step decay; initial lr $=0.002$ $\gamma=0.1$ , milestones [10, 20]	pose noise $(0.2\,\text{m},0.2^{\circ})$
V2X-ViT	60 clean + 10 fine-tune	Multi-step decay; initial lr $=0.001$ $\gamma=0.1$ , milestones [15, 50]	Clean pretraining; fine-tune with pose noise $(0.2\,\text{m},0.2^{\circ})$ and random delay 200–300 ms
TraF-Align baseST	60	One-cycle schedule; peak lr $=10^{-4}$ initial lr $=10^{-5}$ ; ramp-up fraction $=0.4$	random delay 0–400 ms
ERMVP CoST	60	Cosine annealing with warmup warmup lr $=2\times 10^{-4}$ warmup epochs $=10$ minimum lr $=2\times 10^{-5}$	perfect setting throughout

Appendix

Table 4. Method-specific training configurations.

Module	Parameter	Value	Role
Tracking	$k_{\mathrm{hist}}$	3	Number of historical frames used to build short tracklets.
Matching	$d_{\mathrm{gate}}$	4.0	Position gate applied before a candidate pair is considered.
	$\lambda_{\mathrm{vel}},\,\lambda_{\mathrm{graph}}$	1.0, 0.5	Relative weights of the velocity and local-graph terms in Eq. ( 12 ).
	$T_{\mathrm{vel}},\,T_{\mathrm{graph}}$	2.0, 3.0	Normalizers for the velocity and graph terms in Eq. ( 12 ).
	$K,\,w_{\mathrm{edge}},\,w_{\mathrm{pair}}$	3, 1.0, 0.5	Neighbor count and edge/pair signature weights for local layout consistency.
Pose	$N_{\mathrm{trk}}^{\min}$	3	Minimum number of matched trajectories required before running pose refinement.

Appendix

Table 5. Shared structural coefficients used in all experiments.

Group	Parameter	Value	Role
Prior mapping	$T_{q}$	1.5	Cost-to-initial-weight temperature in Eq. ( 13 ).
	$\alpha_{p}$	0.1	Position prior scale in Eq. ( 17 ).
	$\alpha_{v}$	0.03	Velocity prior variance in Eq. ( 17 ).
	$\alpha_{\psi}^{\mathrm{deg}}$	3.0	Yaw prior std factor in Eq. ( 17 ).
Process noise $Q$	$\sigma_{a}$	0.03	Translational process noise in Eq. ( 19 ).
	$\sigma_{\psi,Q}^{\mathrm{deg}}$	4.0	Yaw process noise in Eq. ( 19 ).

Appendix

Table 6. Calibrated Kalman-style uncertainty parameters.

Parameter	Default	Test values	$\Delta$ [email protected]
Distance gate (m)	4	2 / 6	$-0.62$ / $-0.05$
History length (frames)	3	2 / 4	$-0.10$ / $+0.07$
NIS gate (%)	95	90 / 99	$-0.08$ / $-0.07$
Feedback decay	1.5	1.0 / 2.0	$-0.08$ / $-0.04$
Huber penalty	0.5	0.25 / 1.0	$-0.17$ / $-0.29$
Minimum solver support	3	4 / 5	$-0.10$ / $-0.15$

Appendix

Table 7. Hyperparameter sensitivity on V2V4Real under Joint-Hard . The default [email protected] is 67.04.

Panel of Figure 7 — Figure 7. Representative qualitative example of the object-level pipeline in CoAnchor . This example visualizes how CoAnchor progressively improves delayed collaborator alignment through object-level pose refinement, Closed-Loop Posterior Scoring, and temporal propagation. Three vertically arranged bird's-eye-view panels show the same OPV2V scene after initial pose refinement, after closed-loop suppression of unreliable matched pairs, and after propagation to the current ego time. Neighbor boxes become better aligned with ego and ground-truth boxes across the three stages, although some residual motion error remains.

Panel of Figure 8 — Figure 8. Additional robustness curves omitted from the main text. Comparison with state-of-the-art methods on OPV2V and V2V4Real. Each entry is reported as [email protected] / [email protected]. Four line charts complement the robustness curves in the main text. The first two show OPV2V average precision as communication delay increases under fixed pose noise. The last two show V2V4Real average precision as pose noise increases under a fixed communication delay.

[email protected] / [email protected] in OPV2V (100ms fixed)
Method	None	$(0.3\,\text{m},0.3^{\circ})$	$(0.6\,\text{m},0.6^{\circ})$
CoAlign ( Lu et al., 2023 )	95.24 / 80.50	91.61 / 76.39	88.52 / 74.90
TraF-Align (ego-only)	84.49 / 78.57	84.49 / 78.57	84.49 / 78.57
TraF-Align ( Song et al., 2025 )	96.72 / 92.12	89.50 / 73.81	84.10 / 68.95
baseST ( Lu et al., 2023 ; Song et al., 2025 )	96.58 / 91.55	96.26 / 90.06	89.89 / 80.55

Appendix

Table 8. Additional motivation measurements. Method comparison under fixed 100ms communication delay with varying Gaussian pose noise.

Method	All frames	High-motion subset
TraF-Align ( Song et al., 2025 )	58.18 / 35.09	56.35 / 33.50
baseST ( Lu et al., 2023 ; Song et al., 2025 )	61.04 / 37.69	58.83 / 36.01
CoAnchor	67.04 / 44.05	62.76 / 40.63

Appendix

Table 9. Results on the V2V4Real high-motion subset under Joint-Hard . Each entry is [email protected] / [email protected].

Explore similar work

Temporal Residual Bottleneck for Robust Asynchronous Collaborative Perception
Oct 7, 2026Melih Yazgan, Ahmed Abouelazm, J. Marius ZöllnerCooperative Perception Autonomous Driving Perception
FZI Research Center for Information Technology · Karlsruhe Institute of Technology
EgoRefine: Ego-Referenced Predictive Alignment and Trajectory-Conditioned Reliability-Aware Fusion for Asynchronous Collaborative Perception
Sep 29, 2026Lingzhao Kong, Yongsheng Zang, Yu Kang +4Cooperative Perception 3D Object Detection
School of Computer Science and Electronic Engineering, Hunan University, Changsha 410082, China · School of Artificial Intelligence and Robotics, Hunan University, Changsha 410012, China · National Engineering Research Center of Robot Visual Perception and Control Technology, Hunan University, Changsha 410082, China +1
UECP: Uncertainty-Enhanced Collaborative Perception
Jun 22, 2026Kang Yang, Tianci Bu, Peng Wang +3Uncertainty Quantification Autonomous Driving Perception
Renmin University of China, Beijing, China · College of Systems Engineering, National University of Defense Technology, Changsha, China · Intelligent Shipping Center, China Waterborne Transport Research Institute, Beijing, China

Oct 7, 2026, cs.CVJ/K move · Enter open · S save

Temporal Residual Bottleneck for Robust Asynchronous Collaborative Perception

Melih Yazgan, Ahmed Abouelazm, J. Marius Zöllner

FZI Research Center for Information Technology · Karlsruhe Institute of Technology

Cooperative Perception Autonomous Driving Perception Residual Learning

Open paper AskPDF arXiv

Collaborative perception extends the sensing range of autonomous vehicles, but its performance degrades when shared features arrive stale or incomplete. Most latency-robust methods compensate delayed collaborator features through flow-guided alignment or direct feature transport. In this work, we formulate asynchronous collaborative perception as temporal residual prediction. Our Temporal Residual Bottleneck keeps a deterministic pose-warped collaborator feature as a conservative anchor and uses a

Δt

-conditioned xLSTM to extract residual temporal evidence from the available history. A detector-facing residual bottleneck then applies only gated, regularized corrections before ego-side fusion, reducing the risk of overwriting reliable static structure when temporal correspondence is uncertain. Experiments on DAIR-V2X and OPV2V show that our method is especially effective under severe fixed/irregular delays and packet drops. On DAIR-V2X, the reported checkpoint trades a small amount of synchronized peak accuracy for better robustness under stronger communication degradation. Controlled diagnostics further indicate that direct feature transport has oracle headroom but can become unreliable when deployed without accurate correspondence. These results support temporal residual fusion as a practical alternative for asynchronous and incomplete collaborative perception. Code will be publicly released at https://url.fzi.de/8dk38.

Abstract

Explore similar work

Temporal Residual Bottleneck for Robust Asynchronous Collaborative Perception

EgoRefine: Ego-Referenced Predictive Alignment and Trajectory-Conditioned Reliability-Aware Fusion for Asynchronous Collaborative Perception

UECP: Uncertainty-Enhanced Collaborative Perception