基于光视觉信息自适应聚合的船舶自主避碰决策

Autonomous collision avoidance decision-making of ships based on adaptive aggregation of optical visual information

船舶航行环境复杂多变,多模态视觉相机采集的多源视觉信息融合过程繁杂,特征提取不够精细,为此研究光视觉信息自适应聚合的船舶自主避碰决策方法。利用Kinect相机采集多源图像,通过自适应聚合增强环境感知。结合MaskRCNN与ResNet-101,精确提取图像信息中的目标特征,经RPN生成候选区域。特征向量进入FCN与分类回归分支,识别目标并定位,转换坐标后获取运动态势。依据碰撞风险计算和船舶操纵经验,智能输出避碰指令。实验结果表明,该方法在复杂航行环境下展现出良好的避碰决策性能,光视觉信息聚合后,交并比、F1 score分别为92.7%、94.2%。

The navigation environment of ships is complex and ever-changing,and the fusion process of multi-source visual information collected by multimodal visual cameras is complicated.The feature extraction is not precise enough.Therefore,a ship autonomous collision avoidance decision-making method based on adaptive aggregation of optical visual information is studied.Use Kinect camera to capture multi-source images and enhance environmental perception through ad-aptive aggregation.Combining Mask RCNN and ResNet-101,accurately extract target features from image information and generate candidate regions through RPN.The feature vector enters the FCN and classification regression branches,identifies the target and locates it,and obtains the motion situation after coordinate transformation.Based on collision risk calculation and ship maneuvering experience,intelligently output collision avoidance commands.The experimental results show that this method exhibits good collision avoidance decision-making performance in complex navigation environments.After the ag-gregation of optical visual information,the intersection to union ratio and F1 score are 92.7%and 94.2%,respectively.