AGE-Capture: Autonomous, General, and Efficient UAV Capture for High-Fidelity Scene Reconstruction
Organizations: Politecnico di Milano · University of Chinese Academy of Sciences · Zhejiang University
Abstract
High-fidelity scene reconstruction with an unmanned aerial vehicle (UAV) requires more than identifying valuable viewpoints: the system must autonomously acquire informative observations and determine task completion under different sensing configurations and limited flight resources. We present AGE-Capture, an autonomous, sensor-adaptive, and efficient active-capture framework for real UAVs. Building on renderability, the framework uses a lightweight online state to evaluate the utility of candidate views, assess task progress, and guide flight execution. By decoupling geometric exploration from reconstruction observations and establishing a unified camera-ray interface, the framework allows different range sensors and RGB camera models to share a reconstruction-driven acquisition pipeline. Reconstruction-frontier viewpoints characterize residual acquisition value for region screening and autonomous termination; dynamics-aware hierarchical planning combined with asynchronous event-driven capture reduces unnecessary inter-region motion, time spent waiting for plans, and repeated observations. Across ten indoor and outdoor simulation scenes, AGE-Capture achieves a mean peak signal-to-noise ratio (PSNR) -- higher than that of the online baselines while using of the common mission-time budget on average. Its mean peak GPU memory usage is -- lower than that of the same baselines. Experiments with different sensor configurations validate the adaptability of the shared acquisition pipeline, while flights in nine real-world scenes further demonstrate onboard closed-loop operation under practical computational and motion constraints.
Figures & tables
| Protocol item | Online active capture | Offline view placement |
|---|---|---|
| Methods | Ours; ActiveGS; TRACE | Ours; PCP; FPS+Rand. |
| Backend | Incremental 2DGS | Offline 3DGS |
| Budget | elapsed system time | Matched views per scene |
| Stopping | Ours: completion or ; baselines: | Exactly views |
| Shared setup | Initial pose, sensing configuration, planning space, and UAV dynamics | Scene, sensing configuration, camera parameters, and evaluation views |
| Hardware | Ours: RTX 4060 Laptop GPU (8 GB); baselines: A100 SXM4 GPU (64 GB) | A100 SXM4 GPU (64 GB) |
| Scene | Matched Observations | System Time (s) |
|---|---|---|
| Abandoned Warehouse | 339 | 461.0 |
| St Peter’s Church | 520 | 760.7 |
| Barn | 393 | 536.5 |
| Garage Room | 552 | 791.7 |
| Station Arch | 321 | 430.0 |
| Graffiti Alley | 593 | 837.2 |
| PSNR | SSIM | LPIPS | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Scene | ActiveGS | TRACE | Ours | ActiveGS | TRACE | Ours | ActiveGS | TRACE | Ours |
| Abandoned Warehouse | 24.19 | 21.19 | 24.92 | 0.794 | 0.581 | 0.803 | 0.267 | 0.419 | 0.270 |
| St Peter’s Church | 20.62 | 24.02 | 26.00 | 0.686 | 0.764 | 0.802 | 0.413 | 0.341 | 0.327 |
| Barn | 21.29 | 19.32 | 21.70 | 0.660 | 0.620 | 0.672 | 0.319 | 0.368 | 0.328 |
| Garage Room | 24.53 | 23.53 | 26.07 | 0.662 | 0.642 | 0.663 | 0.483 | 0.489 | 0.546 |
| Station Arch | 15.48 | 16.10 | 26.23 | 0.407 | 0.419 | 0.756 | 0.644 | 0.629 | 0.371 |
| PSNR | SSIM | LPIPS | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Scene | FPS+Random | PCP | Ours | FPS+Random | PCP | Ours | FPS+Random | PCP | Ours |
| Abandoned Warehouse | 26.05 | 25.95 | 26.18 | 0.8387 | 0.8331 | 0.8332 | 0.2108 | 0.2336 | 0.2189 |
| St Peter’s Church | 24.41 | 21.93 | 25.22 | 0.8109 | 0.7512 | 0.8315 | 0.3402 | 0.4151 | 0.3164 |
| Barn | 20.40 | 20.61 | 22.03 | 0.6640 | 0.6637 | 0.7186 | 0.3132 | 0.3324 | 0.2725 |
| Garage Room | 26.90 | 24.80 | 26.84 | 0.7251 | 0.6783 | 0.7166 | 0.4424 | 0.5174 | 0.4597 |
| Station Arch | 25.27 | 24.94 | 25.58 | 0.7568 | 0.7413 | 0.7744 | 0.3008 | 0.3140 | 0.2759 |
| Stream | Configuration | Projection / model | Resolution / sampling | Angular coverage | Valid range |
| Range | Depth Map | Pinhole | px | – m | |
| LiDAR | 32-beam spinning | beams | – m | ||
| Pinhole | Pinhole | px | – | ||
| RGB | Fisheye | Double sphere | px | circular FoV | – |
| Panoramic | Equirectangular | px | – | ||
| Pinhole: px. Double sphere: px, , and . | |||||
| Pinhole | Fisheye | Panoramic | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Scene | PSNR | SSIM | LPIPS | Mission Time (s) | PSNR | SSIM | LPIPS | Mission Time (s) | PSNR | SSIM | LPIPS | Mission Time (s) |
| Valve House | 24.19 | 0.636 | 0.431 | 470.5 | 26.51 | 0.731 | 0.400 | 449.5 | 27.04 | 0.714 | 0.389 | 170.3 |
| St Peter’s Church | 21.47 | 0.748 | 0.510 | 760.7 | 24.17 | 0.789 | 0.453 | 481.5 | 25.50 | 0.806 | 0.435 | 333.7 |
| Graffiti Alley | 20.88 | 0.636 | 0.419 | 837.2 | 22.87 | 0.716 | 0.381 | 502.0 | 23.84 | 0.736 | 0.363 | 341.3 |
| Fougères Castle | 23.37 | 0.606 | 0.484 | 168.3 | 23.66 | 0.653 | 0.441 | 124.0 | 25.85 | 0.707 | 0.373 | 73.7 |
| Average | 22.48 | 0.657 | 0.461 | 559.2 | 24.30 | 0.722 | 0.419 | 389.2 | 25.56 | 0.741 | 0.390 | 229.8 |
| Scene | ActiveGS | TRACE | Ours |
|---|---|---|---|
| Abandoned Warehouse | 15.85 | 22.39 | 1.01 |
| St Peter’s Church | 19.77 | 32.09 | 1.58 |
| Barn | 13.63 | 14.25 | 0.96 |
| Garage Room | 16.56 | 28.48 | 1.04 |
| Station Arch | 14.83 | 20.04 | 1.15 |
| Graffiti Alley | 18.96 | 24.27 | 1.38 |
| Range | RGB | PSNR | SSIM | LPIPS | Time (s) |
|---|---|---|---|---|---|
| Pinhole | 23.37 | 0.606 | 0.484 | 168.3 | |
| Depth Map | Fisheye | 23.66 | 0.653 | 0.441 | 124.0 |
| Panoramic | 25.85 | 0.707 | 0.373 | 73.7 | |
| Pinhole | 22.96 | 0.606 | 0.487 | 155.4 | |
| LiDAR | Fisheye | 24.00 | 0.660 | 0.434 | 137.5 |
| Panoramic | 26.05 | 0.702 | 0.377 | 86.4 |
| Component | Specification |
|---|---|
| Platform | |
| Airframe | Carbon-fiber quadrotor |
| Onboard computer | NVIDIA Jetson Orin NX, 16 GB |
| Flight controller | CUAV Nora |
| Battery | 6S, 6000 mAh |
| Sensing and acquisition |
| Setting | Value |
|---|---|
| Motion and control | |
| Speed limit (m/s) | 1.0 |
| Acceleration limit (m/s²) | 0.5 |
| Yaw rate limit (rad/s) | 2.0 |
| Yaw acceleration limit (rad/s²) | 4.0 |
| Command publishing rate (Hz) | 50 |
| Scene | Time (s) | Frames | PSNR | SSIM | LPIPS |
|---|---|---|---|---|---|
| Workshop | 278.9 | 98 | 27.16 | 0.875 | 0.317 |
| Office | 183.8 | 66 | 24.37 | 0.907 | 0.308 |
| Hall 1 | 132.2 | 58 | 28.74 | 0.777 | 0.465 |
| Hall 2 | 138.9 | 49 | 25.39 | 0.771 | 0.440 |
| Garage | 256.4 | 115 | 28.96 | 0.768 | 0.529 |
| Flowerbed | 205.9 | 91 | 21.40 | 0.640 | 0.297 |