摘要
为了提高无人艇在自主航行过程中对水面常见障碍物检测的精度,解决模型参数量较大、模型复杂难以应用于嵌入式设备的问题,提出一种改进的YOLOv3水面常见障碍物检测方法。使用K-means++算法对自建数据集进行聚类得到新的锚框参数,通过添加雨雾噪声的数据增强方法优化模型在复杂天气状况下的障碍物检测能力。针对模型参数量较大问题,使用深度可分离卷积和注意力机制模块重构特征提取网络中的残差结构。为了优化预测框的回归效果,引入SIo U损失函数,将预测框与真实框的方向角度作为损失之一,加快训练速度,提高推理的准确性。通过试验验证了改进后模型参数量缩减了44%,检测精度提高了5.19%,漏检率也有所降低,能有效进行水面障碍物的检测。
In order to improve the detection accuracy of the common obstacles on the water surface during the autonomous navigation of the USV,and to solve the problems that the classical detection model has a large amount of parameters and the model is complex,which is not suitable for embedded devices,an improved method for detecting common obstacles on water surface based on YOLOv3 is proposed.The K-means++algorithm is used to cluster our own dataset to obtain the new anchor boxes parameters,then the obstacle detection ability of the model under complex weather conditions is optimized through data enhancement methods of adding rainy noise and mist-degraded noise to the image.To solve the problem of large amount of model parameters,deep separable convolution and attention module is used to reconstruct the residual structure in the feature extraction network.In order to optimize the regression effect of the bounding box,the SIoU loss function,and taking the direction and angle between the bounding box and the ground truth as one of the losses is introduced to speed up the training and improve the accuracy of reasoning.The experimental results show that the amount of parameters of the model after improvement is reduced by 44%after,and the accuracy of the model is improved by 5.19%,the missed detection rate decreases.It can effectively detect the obstacles on the water surface.
作者
管延敏
汪恭志
余钱程
钟璐阳
虞嘉晨
GUAN Yanmin;WANG Gongzhi;YU Qiancheng;ZHONG Luyang;YU Jiachen(School of Ship and Ocean Engineering,Jiangsu University of Science and Technology,Zhenjiang 212003,Jiangsu,China)
出处
《船舶工程》
CSCD
北大核心
2023年第9期104-113,共10页
Ship Engineering
基金
国家重点研发计划(2022YFE0107000)
国家自然科学基金面上项目(52171259)
关键词
深度学习
目标检测
无人艇
水面障碍物
deep learning
target detection
unmanned craft
surface obstacle
