大尺度海洋环境下AUV的多动态目标跟踪理论与关键技术研究

For the adaptive autonomy of AUV, it is an urgent need to detect the underwater environments accurately. As the basic supporting technology of AUV environment perception, the ability of moving targets detection in dynamic scenes can provide the key technical support for AUV autonomous behaviours, such as path planning, situation assessment and objects avoidance. To focus on the development of autonomous capabilities and aim to improve the perception of underwater environment, the project will centre on the multi-dynamic targets robust tracking theory and key technologies of AUV in large-scale marine environment: The target segmentation theory model is constructed, which the complex marine environment, the platform motion characteristics and the priori shape representation are integrated in.The multi-underwater dynamic target tracking method based on multi-feature description is improved. The deep learning tracking method based on analogue-like sampling and convolutional neural network is studied. A collaborative tracking method of underwater moving target based on multi-cue learning is proposed. It is expected the problem of the continuous tracking of multiple moving targets under the interference of underwater background noise will be solved. In the project, the traditional underwater acoustic detection technology and emerging artificial intelligence technology will be realized, and the key technology research are combined with systematic basic theory perfectly, the achievement will promote the AUV underwater perception technology and application of research progress.

智能无人水下航行器(AUV)的自适应智能发展对水下关联环境的准确感知提出了迫切需求。动态场景中运动目标理解是AUV环境感知的基础支撑技术，可为路径规划、态势评估、危险规避等自主行为提供关键技术保障。本项目围绕AUV自主能力发展需求，以提升AUV水下环境感知能力为目标，开展大尺度海洋环境下基于AUV声视觉的多动态目标跟踪理论和关键技术研究：构建融合复杂海洋环境、平台运动特性以及运动目标先验形状表达和估计为一体的目标分割理论模型；改进基于多特征描述的多水下动态目标跟踪方法；研究基于似物性采样和卷积神经网络的深度学习跟踪方法；提出面向多线索学习的水下运动目标协同跟踪方法。希望通过本项目的实施，解决水下背景噪声干扰下的多运动目标持续跟踪问题，实现传统水声探测技术与新兴人工智能技术、强烈工程发展需求牵引下的关键技术研究与系统性基础理论研究的结合，推动AUV水下感知技术与应用研究取得进步。

基于AUV环境感知能力提升的迫切需求，针对海洋环境中声学成像特有的目标特征点缺失、目标边缘信息点多变以及背景干扰强烈等问题，本项目从水下多运动目标声视觉信息后置处理方法、基于多特征描述的水下多运动目标跟踪方法、基于深度学习跟踪方法等方面入手，按照既定研究方案和技术路线开展相关研究工作，解决了目标区域划分、多目标干扰下的粒子跟踪模型改进、融合信息下的目标模型表述等科学问题，并进行了仿真实验和水池实试验的验证。项目执行过程中，经费严格按照预算执行。在四年的研究过程中，项目组在国际重要刊物上发表论文4篇，其中在国内外SCI检索学术期刊发表论文3篇。申请国家发明专利7项，其中获得授权2项，获省部级奖励1项，完成了项目计划书规定的研究目标评价。