球形视觉模型及全动态场景目标跟踪方法研究

To enhance the effectiveness and performance of mobile imaging devices in scene monitoring and video data analysis, we focus on the data obtained from a group of cameras that distributes on the surface of a sphere, in order to research the computer vision model of spherical surface. Here the formation mechanism of sphere vision, as well as its theoretical problem in applications, is studied by projecting the data onto curve space to implement the task of video processing and analysis. Our aim is to propose and verify a stereo tracking theory and approach to pinpoint the target to be chased in omni-direction nowhere static scene. First, the multi-dimension data obtained from spherical camera array would be transited to a sphere image. We would study the data transitive relations and establish a generalized theory and approach in sphere vision formation and in curve space image processing. Furthermore, since the parasitic motions would impact on the imaging process when the devices move, a method of extracting relative movement from observations and that of parasitic motions restraining will be proposed; meanwhile, a spherical volume vision formation mechanism, an error transition-based method and precision estimation model will be given. And on that basis, we would study the selection mechanism of target chasing clues in the nowhere static scene, and the stereo tracking method of continuous deformation target in curve space as well. Spherical vision is the effective measurement in acting continuously track under the nowhere-static scene monitoring tasks and omni-directional surveillances on continuous deformation target tracking.

为提高成像设备在运动中对全场景进行监视和分析的效率和效果，课题以分布于球形表面的摄像机阵列获得的视频数据为对象，研究球面机器视觉模型，以新的视角- - 曲线空间中的视频处理和分析- - 研究球面机器视觉的形成机制和基础应用环节中的理论问题，以及全动态场景中目标全方位跟踪方法。首先，研究从球形摄像机阵列获得的高维图像集合到球面图像之间的数据优化传递关系，建立球面视觉的形成方法和曲线空间中更具一般性的图像处理理论和方法；进而研究成像设备运动过程中，寄生运动对成像过程的影响，提出从观测运动中相对运动的分离方法和寄生运动的抑制方法，提出球面立体视觉的形成机制、误差传递方法以及精度估计模型；在此基础上研究全动态场景中目标跟踪线索的选择机制和描述方法以及曲线空间中连续形变目标的立体跟踪方法。球面视觉是实现全动态场景监视和连续形变目标全方位不间断跟踪的有效途径。

为提高成像设备在运动中对全场景进行监视和分析的效率和效果，课题以分布于球形表面的摄像机阵列获得的视频数据为对象，研究球面机器视觉模型，通过曲线空间中的图像序列处理和分析，研究球面机器视觉的形成机制和基础应用问题，完成了球形曲面视觉及球面凸分布多目视觉传感器设计；提出一种通过变焦折反射光学系统来协调图像空间分辨率与视场的矛盾关系的思路，考察更为一般化的任意空间曲面成像机理及应用技术，研制了一种主动视觉系统——自由曲面折反射成像装置。该装置可根据任务需求及时地转移注意力焦点，达到一种始终监视全局，又能跟踪目标细节的智能性。基于此开展了大场景中有形变目标检测、跟踪方法研究。一方面，开展了连续变形目标包括视频中的聚集人群目标的异常检测与跟踪研究；另一方面，研究了面向深空探测的太阳系小天体接近段和交会段用于自主导航的目标跟踪方法，即复杂背景中小目标视觉捕获与跟踪，主要用于探测器的相对导航。