Study on Target Recognition and Localization Technology for Underwater Robots Based on Monocular Visual Perception

Abstract Low-cost autonomous underwater vehicles (AUVs) and remotely operated vehicles (ROVs) often rely on a single monocular camera, yet underwater color cast, low contrast, and flow-induced disturbances degrade feature matching and destabilize 3D target localization. This work develops an integrated monocular-vision-based pipeline that combines underwater image enhancement, robust SURF-based feature matching, and geometric localization using a reference-depth image, making it suitable for low-cost and lightweight underwater vehicles. Experimental evaluations show the system achieves an average positioning error of 7.43% with a processing delay of less than 0.041 s per frame in static underwater conditions. Under flow disturbances up to 0.2 m/s, it maintains a maximum error within 13% and stable localization performance. The system also demonstrates adaptability to various target geometries (square, circle, and triangle), with average localization errors of 6.7%, 8.9%, and 11.5%, respectively. Notably, square targets achieve 100% successful localization across all frames. Compared with SIFT and HOG baselines, the SURF-based solution offers a favorable accuracy-efficiency tradeoff, enabling real-time underwater target recognition and spatial localization.