Research map / Detection and segmentation / 3D detection and driving
Drones Help Drones: A Collaborative Framework for Multi-Drone Object Trajectory Prediction and Beyond
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DHD, a multi-drone collaborative trajectory prediction framework with ground-prior BEV generation and sliding-window sparse interaction, plus the simulated Air-Co-Pred dataset built on CARLA.
Problem
Collaborative object trajectory prediction across multiple drones faces inaccurate BEV generation from expansive aerial views (large depth ranges) and excessive inter-drone communication under limited bandwidth.
Approach
Uses the drone's inclined viewing geometry to derive a per-pixel depth upper bound and estimates object height instead of depth (depth = upper bound times (1 - h/H)), and selects features for transmission by sliding-window information volume (feature discrepancy between center and neighbors) with complementary scoring between drones.
Key results
- DHD reaches 61.39% short-range IoU and 50.43% VPQ, versus 56.29% and 44.98% for LSS, and reduces position deviation to 0.82 m (short) from 1.08 m.
- With a 25% transmission ratio, DHD achieves 61.4% short-range IoU versus 62.8% for early collaboration, with the same bandwidth as sparse interaction at a quarter of the fully connected ratio.
- On CoPerception-UAVs collaborative 3D detection, DHD reaches 0.437 mAP versus 0.387 for DVDET and 0.349 for BEVDet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GBG instead of LSS for BEV generation | DHD · Air-Co-Pred short | IoU (%): 56.29 -> 61.39 | Helped |
| Use GBG instead of LSS for BEV generation | DHD · Air-Co-Pred long | IoU (%): 51.76 -> 53.95 | Helped |
| Add ground-prior-based depth estimation (GBG) (ablation) | DHD · Air-Co-Pred short | IoU (%): 57.1 -> 61.8 | Helped |
| Add ground-prior-based depth estimation (GBG) (ablation) | DHD · Air-Co-Pred long | IoU (%): 52.1 -> 54.5 | Helped |
| Add sparse interaction via sliding windows (SISW) (ablation) | DHD · Air-Co-Pred short | IoU (%): 57.1 -> 56.3 | Hurt |
| Add sparse interaction via sliding windows (SISW) (ablation) | DHD · Air-Co-Pred long | IoU (%): 52.1 -> 51.8 | Hurt |
| Add SISW to GBG (ablation) | DHD · Air-Co-Pred short | IoU (%): 61.8 -> 61.4 | Hurt |
| Add SISW to GBG (ablation) | DHD · Air-Co-Pred long | IoU (%): 54.5 -> 54.0 | Hurt |
| Reduce transmission ratio from 1 to 0.25 with SISW (ablation) | DHD · Air-Co-Pred | IoU (%) | No clear effect |
| Remove height estimation from GBG (ablation) | DHD · Air-Co-Pred long | Deviation (m): 1.46 -> 1.67 | Hurt |
| Remove GRU-based temporal fusion (ablation) | UMC · Air-Co-Pred long | IoU (%): 52.3 -> 53.2 | Helped |
| Use DHD rather than no collaboration | DHD · Air-Co-Pred short | IoU (%): 38.1 -> 61.4 | Helped |
Surprising findings
- DVDET's deformable attention improves short-range results but declines in long-range performance relative to vanilla LSS.
- The DHD variant without height estimation still reduces long-range localization error, relying only on ground priors.
Limitations
- Evaluated only in an idealized simulated setting (CARLA) without latency or camera extrinsic noise.
- Air-Co-Pred focuses mainly on the vehicle category due to tiny aerial objects.
- DVDET outperforms DHD on orientation error in 3D detection.
Experimental setup
Datasets: Air-Co-Pred CoPerception-UAVs
Benchmarks: Air-Co-Pred CoPerception-UAVs
Metrics: IoU VPQ precision recall position deviation (m) mAP mATE mASE mAOE
Baselines: LSS DVDET No Collaboration Early Collaboration Late Collaboration V2X-ViT V2VNet Who2com When2com Where2comm UMC BEVDet
Builds on: Lift-Splat-Shoot PowerBEV Where2comm BEVDepth EfficientNet
Closest papers by meaning
- AirCopBench: A Benchmark for Multi-drone Collaborative Embodied Perception and Reasoning (AAAI 2026 · similarity 0.84)
- AGC-Drive: A Large-Scale Dataset for Real-World Aerial-Ground Collaboration in Driving Scenarios (NeurIPS 2025 · similarity 0.83)
- UAV3D: A Large-scale 3D Perception Benchmark for Unmanned Aerial Vehicles (NeurIPS 2024 · similarity 0.83)
- PointSR: Self-Regularized Point Supervision for Drone-View Object Detection (CVPR 2025 · similarity 0.82)
- 3D-aware Select, Expand, and Squeeze Token for Aerial Action Recognition (AAAI 2025 · similarity 0.82)
- JTD-UAV: MLLM-Enhanced Joint Tracking and Description Framework for Anti-UAV Systems (CVPR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.