Answers are grounded in this paper and cite their sources.
Collaborative perception enables super-visual sensing, alleviating the blind spots and limited detection range inherent to single-vehicle perception. In real-world deployments, however, constrained communication bandwidth restricts the amount of information that can be exchanged and introduces inter-vehicle latency. As a result, the delayed features received by the ego vehicle cannot be exploited to fully realize the benefits of collaboration. To address bandwidth limitations, we propose a collaborative perception system designed to make effective use of both the available communication capacity and temporally delayed perception features. Its efficient data communication mechanism can fully compress and transmit full feature-level data. To better exploit temporally misaligned features from neighboring vehicles, we introduce a two-order flow generator grounded in a physical motion model. By leveraging historical observations, it can reliably estimate the current-time feature representation, thereby enhancing collaborative perception performance. In addition, we develop a two-stage training scheme and an end-to-end loss function tailored to the overall framework. Experimental results show that, compared with the traditional approach of transmitting complete feature maps, our method reduces the amount of data transmitted and communication latency by about 4 times. Meanwhile, our proposed two-order flow generator exhibits superior robustness under varying communication delays (100–400 ms) on the V2XSet dataset, maintaining stable performance where competing methods degrade significantly. Finally, we validate the effectiveness of each component through ablation experiments.
No takes yet. Share an insight, caveat, or question.
Hu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: