| With the development of unmanned technology,the study of unmanned surface vehicles has received more and more attention from researchers.Compared with lake and marine environments,unmanned surface vehicles in inland river environments are more closely related to human life.However,the existing research work mainly focuses on the marine environment,and the research and application of unmanned surface vehicles in inland river environment has not been fully developed,one of the main reasons is the inability to provide safe and reliable environment sensing for unmanned surface vehicles systems in the variable and complex environment of inland rivers.Computer vision technology is one of the core technologies for unmanned surface vehicles environment sensing,which is an important way for unmanned surface vehicles to extract surrounding information and also provides information reference for unmanned surface vehicles autonomous navigation decision.In this paper,considering the special environment of inland waters,the method of waters segmentation and target detection for unmanned surface vehicles is proposed,and the deployment and implementation of the algorithm on the unmanned surface vehicles platform is carried out.The main research contents of this paper are as follows:Waters segmentation of unmanned vessels inland rivers.For the problem that the similarity between the shadows of riverbank vegetation,buildings and water surface reflection makes the water boundary blurred,this study proposes a joint dual-task network structure for boundary perception and waters segmentation.Firstly,an attention mechanism module is used to focus on the water information to improve the water segmentation ability,and a feature fusion module is introduced to better integrate the visual features in the decoder with the boundary features.In addition,a boundary-aware module and a boundary loss function are proposed to force the network to focus on the detailed information of the waters boundary.The method is tested using three different types of inland river datasets,and the results show that the method achieves up to 97.19% cross-merge ratio(mIoU)and 0.776 root mean square error(RMSE)of the boundary on the USVInland dataset.It indicates that the method can produce clearer predictions at the watershed boundary and improve the performance of waters segmentation with better generalization performance.Water surface target detection in complex environments.To address the problems of false detection due to vegetation and building reflections on the water surface and small target miss detection,this work proposes an improved YOLOv5 s water surface floating garbage detection method.Firstly,self-calibration convolution is used to expand the perceptual field of the network thus improving the ability of the network to extract features.Secondly,a hybrid feature attention module(CBAM)is used to suppress the interference of noise such as water surface reflection and shadow.Then DeepSort tracking algorithm is added to make the network recognition effect smoother and reduce the target loss problem.In addition,the EIoU loss function is introduced into the network to make the model regress more accurately,thus improving the training effect of the network model.Finally,the experiments are conducted on the test set of the dataset,and 87.4% of mAP as well as 75 FPS are obtained,which verifies the effectiveness and real-time performance of the improved detection algorithm.The implementation of environment sensing system for unmanned surface vehicles.Aiming at the problem of algorithm deployment and implementation on low-cost unmanned surface vehicles,the software and hardware platform of unmanned surface vehicles is built,and the environment sensing system of unmanned surface vehicles is implemented based on JetsonXavierNX and TensorRT,and experiments are conducted on the water surface to provide a reference for practical engineering applications. |
PDF Full Text Request