Abstract
Collaborative perception enables autonomous agents to share sensor data to overcome individual sensing limitations such as occlusions and long-range detection gaps. However, real-world deployments face significant challenges due to pose noise from localization errors and motion blur from dynamic objects. This paper presents MRCNet (Motion-aware Robust Communication Network), a framework designed for robust multi-agent collaborative perception. MRCNet incorporates a Multi-scale Robust Fusion (MRF) strategy to mitigate spatial misalignments and a Motion Enhanced Mechanism (MEM) to leverage temporal context for sharpening dynamic object features. Experimental results on V2XSim, OPV2V, and V2XSet demonstrate that MRCNet achieves state-of-the-art performance and superior robustness against noise while maintaining communication efficiency.
1. Introduction
Autonomous driving relies heavily on accurate perception. While single-agent systems are limited by line-of-sight, collaborative perception allows vehicles and infrastructure to share information, effectively "seeing" through obstacles. Despite its potential, current methods often fail in non-ideal conditions. Specifically, pose noise—originating from GPS or IMU inaccuracies—leads to spatial misalignments when fusing data from different agents. Furthermore, motion blur in sensor data complicates the detection of fast-moving objects. MRCNet addresses these issues by focusing on spatial robustness and temporal enhancement.
2. The MRCNet Framework
The architecture of MRCNet is designed to handle the imperfections of real-world communication and localization. It consists of three primary components: feature extraction, robust fusion, and motion enhancement.
2.1 Multi-scale Robust Fusion (MRF)
To combat pose noise, MRCNet employs the MRF module. Traditional fusion at a single resolution is highly sensitive to coordinate offsets. MRF utilizes Multi-scale Enhanced Aggregation (MEA), which processes features across multiple resolutions. By leveraging high-level semantic features—which are inherently more stable under spatial shifts—the system can provide a "global" anchor to re-align lower-level detailed features. This multi-stage process ensures that even with significant localization errors, the fused representation remains coherent.
2.2 Motion Enhanced Mechanism (MEM)
Dynamic objects often suffer from motion blur or temporary occlusions. The MEM module addresses this by maintaining a historical memory bank of Bird’s-Eye-View (BEV) features. A Gated Motion Unit (GMU) uses deformable convolutions to adaptively align past features with the current frame. This temporal integration allows the network to "sharpen" the perception of moving objects by utilizing information from previous time steps where the object might have been clearer or in a different relative position.
2.3 Communication Efficiency
Bandwidth is a critical constraint in V2X (Vehicle-to-Everything) communication. MRCNet implements a smart feature selection mechanism. Instead of transmitting entire feature maps, it generates a semantic confidence map using channel-wise pooling. This ensures that only the most informative regions for the detection task are shared, reducing communication volume without sacrificing accuracy.
3. Experiments and Evaluation
The performance of MRCNet was validated on three major benchmarks: V2XSim, OPV2V, and V2XSet. V2XSet is particularly important as it includes realistic noise simulations to test robustness.
3.1 Quantitative Results
MRCNet sets a new state-of-the-art (SOTA) on the V2XSim dataset, achieving 85.33% [email protected], outperforming previous models like DiscoNet (83.56%). Similar gains were observed across other datasets, confirming the effectiveness of the proposed modules.
3.2 Robustness to Noise
A key contribution of MRCNet is its resilience to localization errors. In stress tests involving heading errors and positional offsets, MRCNet demonstrated a much slower performance decay compared to existing methods like V2X-ViT and Where2comm.
As shown in the robustness plots, MRCNet maintains a significant lead in Average Precision (AP) even as the standard deviation of heading error increases to 1.0°. This suggests that the multi-scale fusion strategy successfully compensates for coordinate misalignments that typically cripple collaborative systems.
4. Conclusion
MRCNet provides a comprehensive solution for the practical challenges of collaborative perception. By integrating multi-scale spatial fusion and motion-aware temporal enhancement, it achieves high accuracy and robustness in noisy environments. Future work will explore handling communication latency and model heterogeneity to further bridge the gap between simulation and real-world deployment.
Summary updated