Multi-target Detection And Location Methods For Unmanned Surface Vehicles In Typical Ocean Environment

In recent years,Unmanned surface vehicle(USV)has gradually become one of the main forces in ocean engineering with the advantages of relatively low cost,safety and reliability.Its environmental information acquisition mainly depends on visual,infrared,radar and underwater acoustic means.In the actual acquisition of marine environmental information,visual information has higher resolution,contains more edge information,and extracts more effective features.Therefore,this paper takes the"Robot-X" USV with visual perception unit as the research object,mainly studies the detection and location methods of surface targets in typical environments,which can be used to obtain the types and location information of targets or obstacles,and support the subsequent research on risk avoidance and motion planning methods ofUAV.The main contents of this paper are as follows:Firstly,in order to improve the ability of unmanned aerial vehicle to detect sea antenna successfully under poor imaging conditions and poor quality of features,the characteristics and difficulties of sea antenna detection are analyzed,and a sea antenna segmentation data set is constructed by collecting a series of images from unmanned aerial vehicle.Combined with the feature classification technology of neural network theory,a method of sea antenna detection based on full convolution network is proposed.The image pixels are classified by full convolution network,and the classification information is obtained.Then the sea antenna is fitted with the position in the image coordinate system,which solves the instability problem of detection when the linear feature is disturbed.Secondly,aiming at the missing detection phenomenon caused by the small proportion or partial overlap of target pixels in ocean environment,the limitations and optimization directions of existing detection principles are analyzed and summarized,and a multi-scene detection data set of surface targets is constructed according to the working environment and task characteristics of "Robot-X" USV.Combining with the idea of feature fusion,a method of surface target detection based on regional recommendation network is proposed,which improves the recall rate of surface target area recommendation and increases the robustness of detection algorithm to different complex environments.The method has passed the field test verification of real ship.Thirdly,a surface target location strategy based on image information and point cloud information fusion is designed,and information matching criteria are constructed.The types of images,boundary frame information and distance and angle information of lidar are matched and fused,which are applied to "Robot-X" USV and field experiments are carried out to verify the reliability of surface target detection and location methods.