GPS Trajectory Data Enhances YOLOv8 Object Detection Efficiency
A new arXiv preprint (2605.16397) proposes a trajectory-aware adaptive inference method for object detection models in autonomous maritime navigation. The approach integrates GPS trajectory data into a YOLOv8-based detector, introducing an early-exit mechanism that uses motion cues like inter-vessel distances and convergence speed to adjust computational effort. Frames of vessels that are close and converging at high speed are processed with the full model, while others use only a subset of the network architecture. This aims to improve real-time perception efficiency in dynamic maritime environments with large-scale multimodal sensor data.
Key facts
- arXiv preprint 2605.16397 proposes trajectory-aware adaptive inference for object detection.
- Method integrates GPS trajectory data into YOLOv8-based detector.
- Early-exit mechanism uses motion cues such as inter-vessel distances.
- Frames of vessels close and converging at high speed use full model; others use subset of architecture.
- Aims to improve real-time perception efficiency in autonomous maritime navigation.
- Addresses challenges of large-scale multimodal datasets in dynamic environments.
- Object detection and trajectory perception are tightly coupled in maritime navigation.
- Efficiency of object detection models during inference is often overlooked.
Entities
Institutions
- arXiv