抄録
Underactuated autonomous underwater vehicles (AUVs) face significant challenges in benthic monitoring, as their inability to hover and need for continuous forward motion often mismatch with the stationary character of benthic targets. This paper studies how reactive vision-based mission strategies can bridge this gap using a lightweight AUV equipped with real-time YOLOv8-based object detection. We introduce TPM Plus (TPM-P), which augments frame-level detection scores with a bounded confidence state capturing the growth and decay of target belief over time, enabling the vehicle to resume coverage after persistent false positives while retaining the ability to reacquire true targets. In simulation, TPM-P is evaluated against a non-reactive lawnmower baseline and its prior reactive controller (TPM-B). It achieves competitive detection efficiency, detection rate, and detection density while recovering from false positives and maintaining area coverage. We further re-analyze field trials in Rausu, Hokkaido, where TPM-B was deployed with juvenile Red King crabs, quantifying detector precision, recall, and mode transitions; these results highlight high false positive rates and navigation failures, and motivate the design choices embodied in TPM-P. Together, these findings indicate that advanced, reactive monitoring capabilities are feasible on accessible underactuated platforms, while underscoring the need for robust false positive recovery and careful parameter tuning in realistic benthic environments.