水域监控复杂场景下的运动目标检测关键算法研究

Waterway transport plays an important role in comprehensive transportation system of China. There are some difficulties to learn from road traffic video surveillance technology. For example, the camera calibration is so difficult without marking in waterway, such as road traffic. In addition, ship and its reflection in water move together all the time..To resolve the critical security issue of overrun in waterway transport, the intelligent video information processing system is introduced by this project. In the system, the characteristic parameters of moving objects on water like length, width and speed etc. are being monitored in real time by the existing channel camera video, and early warnings are raised timely for overrun cases. Thus the safety can be ensured when the objects pass through bridges or ship locks..The key theoretical areas include: 1) preprocessing of deblurring for surveillance video image in waterway transport to provide high-quality images for subsequent processing; 2) moving foreground detection for water surface area in camera jitter scenes; 3）constructing detection model for reflections of moving objects on water through decision based on multi-feature fusion data, and constructing MDOF (Multi-Degree of Freedom) tracking algorithm model for shift, rotation and scaling to improve the tracking accuracy in case of shelter and deformation; 4) resolving the camera calibration problem in the scenes where it is lack of basis of reference, and improving the calculation accuracy of length and width through considering the characteristic and moving trajectory of the rigid objects.

水路运输因其价廉，承载量大等，在我国综输体系中占有重要地位。然而由于船舶构造尺寸、航行操作不规范，撞桥搁浅、两船相碰等现象时有发生。而借鉴公路交通的视频监控技术，存在种种困难，诸如，航道没有标线，摄像机难以标定；船舶倒影与船行同步，难以区分。本项目针对给水路交通带来极大安全隐患的超限行驶问题，利用已有的航道摄像视频，实时检测水上运动目标的长宽及速度等参数，及时给出超限预警，以保障其在通过桥洞、闸道等特殊地段时的安全。涉及理论问题有：1）水域监控视频图像去模糊预处理问题，为后续图像处理提供更加高质的图像；2）相机抖动场景下水面区域的运动前景检测问题；3）基于多特征融合判决构建水面运动目标倒影检测模型，并构建位移、旋转、缩放多自由度的跟踪算法，提高目标发生相互遮挡和变形的情况下跟踪准确率；4）信息极度匮乏场景中摄像机的标定问题，同时结合运动目标的特性及其运行轨迹，改进长宽尺寸计算的准确度。

水路运输因其价廉，承载量大等，在我国综输体系中占有重要地位。然而由于船舶构造尺寸、航行操作不规范，撞桥搁浅、两船相碰等现象时有发生。课题针对给水路交通带来极大安全隐患的超限行驶问题，利用已有的航道摄像视频，实时检测水上运动目标的长宽及速度等参数，及时给出超限预警，以保障其在通过桥洞、闸道等特殊地段时的安全。涉及工作如下：1）水域监控视频图像去模糊预处理问题，为后续图像处理提供更加高质的图像；2）相机抖动场景下水面区域的运动前景检测；3）融合区域纹理梯度的船舶阴影视频去除算法，提高船舶检测精度；4）研究抑制船尾拖纹的船舶显著性视频检测方法，消除船尾拖纹的影响；5）结合深度学习检测框架完成复杂场景下的船舶检测；6）信息极度匮乏场景中结合刚性运动目标特征实现交互式标定，同时结合其运行轨迹，提高长宽尺寸计算的准确度。课题研究成果已构建应用演示系统，并应用于省市级成果转换项目。