计算机应用 ›› 2019, Vol. 39 ›› Issue (6): 1663-1668.DOI: 10.11772/j.issn.1001-9081.2018102190

• 人工智能 • 上一篇    下一篇

基于Darknet网络和YOLOv3算法的船舶跟踪识别

刘博, 王胜正, 赵建森, 李明峰   

  1. 上海海事大学 商船学院, 上海 201306
  • 收稿日期:2018-10-31 修回日期:2019-01-11 发布日期:2019-06-17 出版日期:2019-06-10
  • 通讯作者: 王胜正
  • 作者简介:刘博(1992-),男,安徽阜阳人,硕士研究生,主要研究方向:船舶智能航行、计算机视觉;王胜正(1976-),男,湖南双峰人,教授,博士生导师,博士,主要研究方向:航海仿真、智能船舶航行、大数据、机器学习;赵建森(1983-),男,黑龙江海林人,副教授,博士,主要研究方向:智能通信、微波与天线;李明峰(1994-),男,四川遂宁人,硕士研究生,主要研究方向:船舶智能航行、强化学习。
  • 基金资助:
    国家自然科学基金资助项目(51379121,61304230);上海市曙光人才计划项目(15SG44)。

Ship tracking and recognition based on Darknet network and YOLOv3 algorithm

LIU Bo, WANG Shengzheng, ZHAO Jiansen, LI Mingfeng   

  1. Merchant Marine College, Shanghai Maritime University, Shanghai 201306, China
  • Received:2018-10-31 Revised:2019-01-11 Online:2019-06-17 Published:2019-06-10
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (51379121, 61304230), the "Dawn" Program of Shanghai Education Commission (15SG44).

摘要: 针对我国沿海与内陆水域区域视频监控处理存在实际利用率低、误差率大、无识别能力、需人工参与等问题,提出基于Darknet网络模型结合YOLOv3算法的船舶跟踪识别方法实现船舶的跟踪并实时检测识别船舶类型,解决了重要监测水域船舶跟踪与识别问题。该方法提出的Darknet网络引入了残差网络的思想,采用跨层跳跃连接方式以增加网络深度,构建船舶深度特征矩阵提取高级船舶特征进行组合学习,得到船舶特征图。在此基础上,引入YOLOv3算法实现基于图像的全局信息进行目标预测,将目标区域预测和目标类别预测整合于单个神经网络模型中。加入惩罚机制来提高帧序列间的船舶特征差异。通过逻辑回归层作二分类预测,实现在准确率较高的情况下快速进行目标跟踪与识别。实验结果表明,提出的算法在30 frame/s的情况下,平均识别精度达到89.5%,与传统以及深度学习算法相比,不仅具有更好的实时性、准确性,对各种环境变化具有较好的鲁棒性,而且可以识别多种船舶的类型及其重要部位。

关键词: 海上交通, 船舶监测, 船舶跟踪, 船舶类型识别, Darknet网络, YOLOv3算法

Abstract: Aiming at the problems of low utilization rate, high error rate, no recognition ability and manual participation in video surveillance processing in coastal and inland waters of China, a new ship tracking and recognition method based on Darknet network model and YOLOv3 algorithm was proposed to realize ship tracking and real-time detection and recognition of ship types, solving the problem of ship tracking and recognition in important monitored waters. In the Darknet network of the proposed method, the idea of residual network was introduced, the cross-layer jump connection was used to increase the depth of the network, and the ship depth feature matrix was constructed to extract advanced ship features for combination learning and obtaining the ship feature map. On the above basis, YOLOv3 algorithm was introduced to realize target prediction based on image global information, and target region prediction and target class prediction were integrated into a single neural network model. Punishment mechanism was added to improve the ship feature difference between frames. By using logistic regression layer for binary classification prediction, target tracking and recognition was able to be realized quickly with high accuracy. The experimental results show that, the proposed algorithm achieves an average recognition accuracy of 89.5% with the speed of 30 frame/s; compared with traditional and deep learning algorithms, it not only has better real-time performance and accuracy, but also has better robustness to various environmental changes, and can recognize the types and important parts of various ships.

Key words: sea traffic, ship surveillance, ship tracking, ship type recognition, Darknet network, YOLOv3 algorithm

中图分类号: