基于RetinaNet的水下机器人目标检测

Target detection of underwater vehicle based on RetinaNet

为解决水下机器人目标检测效率低且环境适应力差的问题,提出一种基于改进的Retina网络水下机器人目标检测方法。采用Dense Net替代Res Net构建骨干网络,使用卷积层堆叠代替原始单次卷积操作,减轻网络重量。以海参为典型研究目标,实验结果表明,所提方法能够正常运行在水下机器人上,相比之前基于机器视觉的目标检测方法,检测精度提高约23%,运行速度提高约17%,在运行速度与检测精度上优于一些基于卷积神经网络的常见目标检测算法。

To solve the problems of low efficiency and poor adaptability of underwater vehicle target detection,an underwater vehicle target detection method based on improved RetinaNet was presented.DenseNet was used to replace ResNet to build the backbone network.Convolution layer stack was used to replace the original single convolution operation to reduce the weight of net.Taking sea cucumber as a typical research object,the experimental results show that the proposed method can operate normally on the underwater robot.Compared with the previous target detection methods based on machine vision,the detection accuracy is improved by about 23%,and the operation speed is improved by about 17%.The operation speed and detection accuracy are better than that of some common target detection algorithms based on convolution neural network.