USV Obstacle Detection via Anomaly Detection
Overview
Pipeline for obstacle perception for unmanned surface vehicles (USVs) in limited geographic areas (rivers, lakes, sea near-shore). It is designed for rapid prototyping and environmental surveillance when collecting large labelled obstacle datasets is impractical. The default offering is offline, batch inference on user-provided image/video sets, producing obstacle maps and overlays for evaluation and integration testing. The pipeline is based on the cascaded datasets method: (1) modality adaptation (e.g., RGB vs LWIR/thermal), (2) environment adaptation learned from obstacle-free runs (one-class anomaly detection), and (3) sensitivity tuning with a small pixel-annotated set to set thresholds. It runs as GPU jobs under Slurm, supports standard anomaly detectors (e.g., PaDiM/CS-Flow) and optional water-edge masking.
Outcome
Users receive per-frame obstacle probability maps, binary masks, QC overlays, and summary reports, plus an optional calibrated model artefact. Optional add-on: rapid site adaptation using the user's obstacle-free data and a small tuning set. Experimental add-on: online inference via reserved GPU workers (best-effort; not safety-certified).
Service Card
| Category | AI products |
| Domain | Maritime and Inland Waterway Autonomy, Environmental Monitoring/Surveillance |
| Vertical | Green Transition |
| Target users | USV Developers/Integrators, Robotics SMEs, R&D Teams, Environmental Monitoring Stakeholders, Research Users |
| Service type | Pipeline |
| Outputs | Per-frame obstacle probability maps, binary masks, overlays, JSON/CSV summary; optional adapted model artefact (checkpoint/config). |
| Requirements | Batch mode is the default (Slurm queue; no latency guarantee). Best results require obstacle-free runs from the target area and (optionally) a small tuning set to calibrate sensitivity. Performance may degrade under strong condition shifts (weather/illumination/season); site adaptation mitigates this. Online mode is experimental, requires reserved resources and network integration, and is not safety-certified. |
| Limitations | |
| Availability | Q3/2027 (batch pilot); online mode by arrangement following that date |
| Licensing | TBD (service implementation: provider-managed; dataset/licenses depend on chosen backends) |
| Visual content | |
| Visual caption |