Bandwidth Efficient Collaborative Perception for Connected Vehicles

204 concepts165 papers1 proposal

Collaborative perception allows autonomous vehicles (AVs) and roadside infrastructure to "pool" their sensor data, effectively seeing through obstructions and extending their sensing range far beyond the limits of a single agent. However, a fundamental bottleneck exists: raw sensor feeds, such as high-resolution LiDAR point clouds or 4K video, generate massive volumes of data that far exceed the capacity of standard Vehicle-to-Everything (V2X) communication networks.

To bridge this gap, researchers have moved away from sharing raw data (early fusion) or simple detection results (late fusion). Instead, the field has converged on intermediate fusion, where agents share compressed, high-value neural features. The goal is to maximize perception accuracy while minimizing the "communication footprint."

Selective Communication via Spatial Importance

The most intuitive way to save bandwidth is to avoid sending redundant information. In a typical driving scene, much of the environment—like the empty sky or a well-lit road surface already seen by all agents—is low-value.

Hu et al. proposed Where2comm, which uses "spatial confidence maps" to identify which areas of a Bird’s Eye View (BEV) feature map are most critical for detection. By only transmitting these high-confidence regions, the authors achieved a 100,000× reduction in communication volume. Similarly, Wang et al. introduced RiSe, which prioritizes data based on "risk-intent." Instead of asking what the infrastructure sees, the system asks what the vehicle needs to know to navigate safely, filtering out irrelevant background data.

Other approaches use uncertainty as a filter. Cheng et al. developed GevBEV, which employs evidential deep learning to quantify spatial uncertainty. Agents only share information about areas where their own "evidence" is low, reducing data sharing by 87% without sacrificing safety.

Compressing Features into Discrete Codes

Even when sharing only the "important" parts of a feature map, the data remains high-dimensional. To further compress these features, researchers have turned to neural quantization and sparse architectures.

Hu et al. introduced CodeFilling, a framework that treats collaborative perception like a jigsaw puzzle. Instead of sending floating-point feature vectors, agents transmit compact "integer codes" from a shared dictionary (a codebook). The receiving agent uses these codes to "fill the gaps" in its own knowledge, slashing costs by over 1,000x.

Sparse transformers also play a key role. Xu et al. proposed CoBEVT, which utilizes a Fused Axial Attention (FAX) mechanism. This architecture allows vehicles to efficiently attend to sparse, relevant features from multiple cameras and neighbors, preserving critical spatial details while ignoring the "noise" of empty space.

Moving Toward Abstract Representations

A more radical shift involves moving away from spatial feature maps entirely, favoring more abstract, low-bandwidth representations like objects, graphs, or even language.

  • Object-Level Queries: Wang et al. proposed CoopDETR, which transmits "object queries" rather than dense maps. By sharing only the learned embeddings of potential detected objects, the authors reduced transmission costs to 1/782 of previous methods.
  • Spatiotemporal Graphs: Gao et al. represented sequences of observations as graphs, achieving a 100x reduction in data size by capturing the relationships between moving objects over time rather than raw pixels.
  • Natural Language: Gao introduced LangCoop, which uses natural language "LangPacks" as the communication medium. This not only reduces bandwidth by 96% but also allows vehicles with entirely different AI architectures to collaborate using a shared human-readable interface.

Mitigating Latency and Asynchrony

Bandwidth constraints often lead to network congestion, which in turn causes communication delays. In a fast-moving traffic environment, a feature map that is 100ms old is no longer spatially aligned with the current world.

To solve this, Wei et al. developed CoBEVFlow, which uses "BEV flow" (motion vectors) to warp historical features into the present time. Similarly, Yu proposed FFNet, which transmits "feature flows"—intermediate representations combined with their rate of change. These methods allow the receiving vehicle to predict where an object should be now, even if the data arrived late, ensuring that collaborative perception remains robust despite the inherent limitations of wireless networks.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 162 papers
  2. Version 1 · 2026-10-04 · 230 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_bandwidth_constrained_collaborative_perc_c650489b,
  title = {bandwidth-constrained collaborative perception},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/bandwidth-constrained-collaborative-perception/rdr_c650489bb2694df58ef35e8def4189c7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 2, frozen text):

@misc{lacuna_bandwidth_constrained_collaborative_perc_c650489b_v2,
  title = {bandwidth-constrained collaborative perception},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/bandwidth-constrained-collaborative-perception/rdr_c650489bb2694df58ef35e8def4189c7/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake