声混叠情况下的水下运动目标图像跟踪技术研究

Taking sonar as the center, forward-looking sonar lacks the recognition capacity for different objects that at the same range and the same horizontal angle, but different vertical angle. When the target moves to the spatial location that satisfied the conditions mentioned above, the echoes of the target and the obstacle are overlapped, which is called the acoustic aliasing phenomenon. Then the echo of the target is cut down or even completely submerged. As a result, the gray, texture, and the statistics information in the sonar image are lost abruptly. Thus, the acoustic aliasing becomes the bottleneck of underwater target tracking. This study focuses on underwater moving target image tracking for two normal operation pattern, namely smooth inspect and looking down. Geography information is still credible if there was no acoustic aliasing, so in the aspect of target contour iteration between frames, the strategy of controlling the inner particles using the shape prior, and the feasibility of mapping the distance between particles and the shape prior to particles weight are discussed. The moving information of the interested target and the energy flow are used to judge the happening and disappearing of the acoustic aliasing. The tracking models are dynamically changed, so as to estimate the moving status of the underwater target. This study aims to improve the Intellectualization of underwater robot vision, and gain the technique for nearshore underwater security and detection.

以声纳为中心，对于处在相同距离、相同水平角度、不同垂直角度的相似物体，前视声纳缺少对它们的分辨能力，当目标运动至与障碍物的空间相对位置满足该关系时，它们的回波叠加在一起，产生声混叠现象，目标回波特征被削减或被淹没，图像中灰度、纹理及统计等信息的可信度骤降，声混叠成为水下目标跟踪技术的瓶颈。本立项针对前视声纳平视与俯视两种常规作业模式，研究水下运动目标图像跟踪技术。鉴于尚未发生声混叠时、几何信息的目标描述能力尚有可信度，拟在帧间目标轮廓迭代方面，主要探讨用形状先验所决定的边界规范粒子重采样的策略、以及将粒子与该决定边界的距离信息映射为粒子权重的可行性。声混叠发生过程中，拟利用目标的运动信息和能量流判断声混叠的发生和消退，动态调整跟踪模型，维持对水下感兴趣目标的运动状态的估计，并通过水池实验数据验证算法的跟踪性能。本研究旨在提高水下机器视觉智能化，为近岸水下安防和水下探测提供技术储备。

项目针对二维前视声纳垂向多目标分辨能力弱的问题，提出利用目标的运动信息和能量流判断声混叠的发生和消退的智能判断机制，采用动态调整水下运动目标跟踪模型的思路，解决目标声回波特征被削减或被淹没时维持对水下目标运动状态估计的难题，实验结果表明，针对所设计的声混叠实验场景，单运动目标的序列跟踪平均丢失率不高于6%，所提出技术具有维持水下运动目标状态估计能力，可以有效提取水下运动目标的运动信息，具有较高的跟踪准确率和较好的跟踪鲁棒性。.项目形成《声混叠情况下的水下运动目标图像跟踪技术研究》技术报告1份。发表学术论文5篇，其中SCI论文3篇、EI论文1篇、EI待检论文1篇。获得中国发明专利授权1项、申请国内发明专利1项、计算机软件著作权登记1项。按照各年度研究计划有序开展研究工作，完成了既定研究内容，达到研究目标，为关键科学问题提供了解决思路，取得了预期研究结果。研究为提高水下机器视觉智能判断程度提供了思路，技术可应用于水下近岸安防，如水下蛙人、水下入侵潜器等的探测。