SonarVoxNet: Diver Detection in 3D Bounding Box using 3D Sonar
This preprint introduces SonarVoxNet, a 3D sonar detector that predicts full 9-degree-of-freedom oriented bounding boxes for divers, plus Diver3D, a public 37,010-frame dataset with full 3D orientation labels. With augmentation and causal refinement, SonarVoxNet reaches 0.774 [email protected] and 0.719 SDS on the held-out test set, far above LiDAR-derived baselines that top out at 0.374 [email protected]. The work shows that pitch and roll of a freely swimming diver can be recovered from sparse acoustic returns, which matters for AUV diver-robot interaction.
Can 3D sonar returns alone support detecting a diver as a full 9-degree-of-freedom oriented bounding box, including pitch and roll?
AUVs that assist human divers need to track not just a diver's 3D position but full-body orientation; optical cameras fail in turbid water, and widely used forward-looking sonar collapses elevation so it cannot estimate pitch and roll. Commercial 3D sonar preserves elevation but produces sparse, noisy returns, and existing point-cloud detectors are designed for dense LiDAR and for upright targets that rotate only in yaw.
Prior underwater 3D sonar detection was limited to generic objects, such as SonarPoint, rather than divers. Prior human detection from LiDAR used voxel-based or point-based backbones such as VoxelNet, SECOND, PointPillars, CenterPoint, and VoxelNeXt, with yaw-only anchor-based or center-based heads. Those heads cannot represent out-of-plane rotation, so they may localize a diver but fail to capture body orientation.
SonarVoxNet uses a sparse voxel encoder and an anchor-free center head, but replaces the yaw-only angle with a continuous 6D rotation parameterization mapped to SO(3). The authors also release Diver3D, collected at Blue Grotto with 36 scenes, 37,010 frames, and 34,568 diver instances. They add an orientation-preserving augmentation curriculum and a causal inference-time refinement stage that uses interacting multiple-model filters for translation and rotation and a shrink-only size update.
On the Diver3D test set, SonarVoxNet without augmentation attains 0.515 [email protected] and 0.233 [email protected], versus 0.126, 0.315, 0.190, 0.324, and 0.374 [email protected] for VoxelNet, SECOND, PointPillars, CenterPoint, and VoxelNeXt, respectively. With augmentation these rise to 0.703 [email protected] and 0.359 [email protected], a relative gain of +37% and +54% over no augmentation. Adding causal refinement yields 0.774 [email protected], 0.467 [email protected], 0.719 SDS, 43.7° full-3D orientation error, and 18.5° tilt error. Ablations show the yaw-only to full-SO(3) transition is the dominant gain: [email protected] rises from 0.328 to 0.538 on a dense backbone and from 0.319 to 0.515 on sparse, while 3D orientation error falls from 70.8° to 51.3° and from 66.8° to 48.5°. Dense and sparse full-SO(3) configs are within noise on SDS, 0.576 vs 0.574. Dataset statistics: 27.6% of frames have no diver, 55.5% have one; diver tilt is standing (<15°) in 34.8% of instances, lying (75–105°) in 30.6%, and intermediate in the remaining 34.5%; no-diver ROC-AUC is 0.614.
The authors state that the z-axis is coarsely compressed before bird’s-eye-view projection, positive examples are scarce at 1.3 divers per non-empty frame, full-rotation labels for a non-rigid diver can be ambiguous, and the pipeline does not yet model acoustic sensor physics or a diver-specific prior. Evident limitations include evaluation on one natural cave-diving site with one class and one sensor type, no reported closed-loop AUV deployment, and a refinement stage that only adjusts already-detected boxes so detection recall is unchanged.
The dataset and benchmark are useful now for AUV manufacturers, underwater inspection and service robotics firms, marine autonomy suppliers, and search-and-rescue or defense operators that need diver detection for human-robot interaction. Product deployment is likely 3+ years away because the approach would need to run in real time on embedded underwater hardware, be validated across varied sites and sensor models, and likely be fused with other sensors for robust field operation.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.