Key Technologies Research On Surface Target Perception For Intelligent Unmanned Surface Vehicles

Environmental perception is a prerequisite for autonomous operation of intelligent unmanned surface vehicles(USVs),and target perception has become one of the research hotspots in this field.In recent years,the rapid development of sensor technology has provided data support for target perception research and application of USVs.However,there are still challenges in effectively perceiving targets using sensor data in complex weather and lighting conditions.On the one hand,a single sensor is limited by its characteristics and environmental factors,unable to achieve multi-dimensional target perception and lacks robustness in complex scenarios.On the other hand,fusion perception algorithms based on multiple sensors can obtain multi-modal information of targets,but require appropriate fusion strategies to process data from multiple sensors.However,current research on target perception based on sensor fusion often focuses on single target detection tasks and does not make full use of various sensors for multi-dimensional information perception,such as semantics and localization.To address these issues,this study focuses on the target perception technology and platform application of USVs in the context of water surface garbage collection tasks,with the main research contributions as follows:(1)A multi-category,multi-scenario water surface floating garbage dataset named Lake Litter-I is created from the perspective of USVs in a real environment to address the problem of limited categories and scenarios in the public water surface garbage dataset Flo W-Img,providing data support for algorithm validation and related research.Moreover,the inaccurate labels in the Flo W-Img dataset are revised to ensure the accuracy of dataset annotations.(2)Aiming at the small target and strong interference in water surface garbage perception tasks,this paper proposes an improved target perception algorithm FE-YOLOv5 n.A channel rearrangement feature enhancement module is designed to extract more expressive features,and a feature focusing module replaces the simple connections in the original network neck to build a feature focusing fusion pyramid,reducing interference from redundant features.The proposed FE-YOLOv5 n achieves m AP of 84.2% and 85% on Flo W-Img and Lake Litter-I datasets,respectively,improving the baseline algorithm by 1.8%and 1.7%.(3)To address the issue of multi-dimensional perception limitations in visual sensor-based target perception algorithms,this paper proposes a multisensor fusion-based target perception method.The proposed method combines FE-YOLOv5 n and point cloud target extraction methods,introducing pixel expansion in the information fusion process to achieve multi-dimensional target perception and enhance the algorithm’s anti-interference capability.The proposed method achieves F1 scores and m AP of 89.7% and 90.5% on the Flo WRI dataset,with root mean square errors of 0.0392 m and 0.0599 m for target distance perception and localization,respectively.(4)To validate the practical detection performance of the proposed target perception algorithm in real-world scenarios,this study analyzes the design requirements of the USV platform,completes the construction of the USV platform and the design of the shore-based system,and deploys the proposed algorithm on the constructed platform.Joint debugging demonstrates the effectiveness of the proposed target perception method.