A Lightweight Deep Learning-Based Fast Detection Algorithm for Waterway Navigation

To address the challenges of waterway detection for unmanned marine equipment in complex aquatic environments within marine ranching, this paper proposes a lightweight deep learning-based fast detection algorithm. By integrating multi-modal image enhancement techniques, the method effectively mitigates issues caused by light reflection, wave interference, and blurred waterway boundaries. An innovatively designed prior-guided ROI row anchor network, combined with a dynamic weight allocation strategy, achieves continuity modeling of waterways through sparse sampling. A multi-constrained loss function significantly enhances detection robustness under harsh environmental conditions. Experimental results demonstrate that the proposed algorithm processes 1080p-resolution frames in 33ms per frame, outperforming the UFLD method with a 52% improvement in F1-score under extreme lighting and submerged waterway scenarios, alongside a 57% reduction in weighted average pixel error. Integrated with a binocular depth camera-based real-time ranging module, the system achieves high-precision distance perception with average errors of 5.6 cm (favorable conditions) and 11.7 cm (adverse conditions) within 5-20m ranges. This provides robust technical support for unmanned surface vehicles in obstacle avoidance, navigation, and precision operations.