Deep Learning based Underwater Object Detection and Recognition for Multi-Robot Systems

This paper introduces an innovative method for underwater object detection and recognition using a deep learning framework and autonomous underwater vehicles. The proposed method uses a deep learning framework with a network of underwater mobile robots operating within three-dimensional environments. To navigate these environments effectively, a random decentralized algorithm guides autonomous underwater vehicles (AUVs) in detecting objects within confined three-dimensional spaces. The deployment strategy involves utilizing the vertices of a cubic grid to optimize the positioning of AUVs across the search area. To enhance the object detection capabilities of mobile underwater robots within these environments and address the complexities of similar objects concealed in their surroundings, deep learning based on the YOLOv۶ framework is refined to achieve improved performance and efficiency metrics. Specifically, the proposed Camouflaged Sea Edge and Attention Module (CSEA) is integrated into the YOLO framework, resulting in a substantial (۸%) enhancement in mean average precision (mAP) performance, with a negligible impact (approximately ۳%) on efficiency when evaluated on the publicly available TrashCan benchmark dataset. Additionally, to accelerate the training convergence and to better detect small objects within the TrashCan dataset, a distinct loss function based on Distance Intersection over Union (DIoU) is employed in the final proposed model. Several ablation studies are conducted to compare various YOLO models and evaluate the performance of the proposed CSEA module on the TrashCan dataset.

کلیدواژه ها:

mobile robots ، AUVs ، Distributed Control ، deep learning ، Object Detection