基于深度学习的海上目标快速检测方法

在智能海洋船艇自主导航和避障过程中,对周边环境的感知能力非常重要。本文针对智能船艇对环境感知的速度和精度需求,提出了基于深度学习的海上目标快速检测方法,并构建了以船舶、浮标和岛屿为检测目标的海上目标数据集。本方法通过提取目标图像不同尺度特征语义信息,应用多框检测器实现对目标的分类和定位,采用非极大值抑制算法筛选最优结果,实现了海上目标快速高精度检测。实验结果表明,本方法在自建数据集上取得了83.3%的识别正确率,单帧耗时9.8ms, 在相同实验条件下,正确率和单帧耗时均优于现有同类快速目标检测方法,证明了该方法可以满足智能船艇对环境感知的速度和精度需求。

A Fast Maritime Target Detection Method Based On Deep Learning

In the process of autonomous navigation and obstacle avoidance of smart ships, it is very important to perceive the surrounding environment. The speed and accuracy requirements of the smart ship"s perception of the environment is focused in this paper. A fast detection method for maritime targets based on deep learning is implemented, and a data set of marine targets with ships, buoys and islands as detection targets is constructed. By extracting the semantic information of the target image at different scales, the multi-box detector is used to achieve the classification and position of the object, and the non-maximum suppression method is used to retain the best results, fast and high precision detection of marine targets is realized. The experimental results showed that with less computing resources, 83.3% recognition accuracy is achieved on the self-built data set. Each frame costs 9.8ms, the accuracy and detection speed are superior to the existing similar detection methods. It is also shown that the proposed model can effectively complete the requirement of smart ship"s perception of the environment.