SBIR Phase II: Auto-Tracking Using Trailing Templates and Skeletal Guides

这项小型企业创新研究（SBIR）第二阶段项目继续研究和开发，旨在证明在不受约束的环境中对动物和人类的运动进行自动视频跟踪的可行性。第一阶段的研究成功地将低级智能设计成预测搜索算法，能够将其在后续图像中的正确位置搜索限制在系统预测的特定小区域内。目标是创建一个软件系统，简单的用户可以轻松操作，通过一系列视频图像快速准确地跟踪移动动物或人类的多个点或区域。这种跟踪可以在背景杂乱和间歇性遮挡的情况下完成，并且不需要将任何区分标记附加到主题上。在第一阶段，设计了一个用户界面，允许用户通过选择闭合多边形的顶点和连接的旋转点来选择要用指向设备（鼠标）跟踪的“骨架模板”。通过感知系统的运动方向和速度，基于模型的跟踪算法告诉搜索机制，它应该在下一张图像中寻找哪里，以匹配从模板的先前位置和方向派生的“跟踪模板”。在第二阶段，将引入更复杂的建模和预测算法，包括对构建模型的监督学习，以及在视频加载时构建的从粗到细的金字塔尺度空间，这将提高跟踪算法的速度和效率，并提高基于模型的方法的鲁棒性。与此同时，用户界面将被重新定义，以改善“外观和感觉”，并赋予其更直观的结构。该软件的应用程序有现成的市场需求。目前商用的生物运动跟踪技术要求在物体的关键位置放置侵入式控制目标。跟踪和描述一般生物运动的商业需求将被利用，包括动物行为分析工具，以及预测和提高运动员的运动效率。此外，这项技术还应用于诊断和医学/健康应用、监测以及其他用途，从美国宇航局的太空研究到人体工程学设计，再到乐器的指法。

This Small Business Innovation Research (SBIR) Phase II project continues research and development aimed at demonstrating the feasibility for automatic video tracking of the motion of animals and humans in unconstrained environments. The Phase I study succeeded by designing low-level intelligence into predictive search algorithms that were able to confine their search for the correct position in a succeeding image to specific, small regions predicted by the system. The objective is to create a software system, easily operable by an unsophisticated user that can quickly and accurately track multiple points or regions of a moving animal or human through a sequence of video images. This tracking can be done despite background clutter and intermittent occlusion, and without attaching any distinguishing markers to the subject. In Phase I, a user interface was designed that allowed the user to choose a 'skeletal template' to be tracked with a pointing device (a mouse) by selecting vertices of closed polygons and connected rotation points. By sensing the direction and speed of motion of the system, the model-based tracking algorithm told the search mechanism where it should look in the next image to match a 'trailing template' derived from previous locations and orientations of the template. In Phase II, more sophisticated modeling and prediction algorithms, including supervised learning of constructed models, and a pyramided coarse-to-fine scale-space, constructed at video load time, will be brought to bear that will increase speed and efficiency of the tracking algorithm and improve the robustness of the model-based approach. At the same time, the user interface will be redefined to improve the 'look and feel' and give it a more intuitive structure. Applications for this software have a ready market demand. Present commercial tracking technology of biological motion requires the placement of intrusive control targets at critical positions on the subject. The commercial need for tracking and characterizing general biological motion will be exploited, including tools for animal behavior analysis, and predicting and improving motion efficiency in athletes. In addition, this technology has applications in diagnostics and medicine/health applications, surveillance, and other uses ranging from NASA's space research, to ergonomic design, to the fingering of musical instruments.