Finding Skeletal Structures of Living Entities and Articulated Objects using Unsupervised Learning on Videos

从图像中提取有用的特征是计算机视觉中必不可少的。许多算法依赖于从图像中提取重要元素，以便对图像包含的内容进行分类。这些年来，特征提取变得更加复杂，在识别图像和无缝构建图像时具有更高的准确度。一个尚未被充分利用的功能是对象在视频中移动的方式，因为大多数现有的研究都只考虑以人类为对象，但它有可能适用于任何有关节点的物体。大多数现有的寻找骨骼结构的研究只集中在寻找人类骨骼上，并且可以假设人类骨骼的形状因此，能够使用监督学习来将形状与已经提供的视频相匹配。这项研究的重点是结合使用计算机视觉技术，以找到任何具有关节的物体的近似骨骼结构，重点是所有类型的动物。使用无监督学习的原因是，它将允许将其应用于任何物种的动物，即使结构差异很大，因此无法对动物身体部位的形状进行假设。这种方法的应用广泛而多样，涵盖了许多跨学科领域，包括：生物图像分析、各种形式的遥感数据分析、医学图像分析（例如，检测身体部位，并分割身体内的连接结构（例如，心脏和肺）），自动驾驶车辆应用（特别是用于在视频片段中检测和建模铰接车辆，其在例如自动驾驶汽车中具有主要应用）。

Extracting useful features from images is essential in computer vision. Many algorithms rely on extracting important elements from an image in order to classify what the image contains. Across the years, feature extraction has become more intricate, allowing higher degrees of accuracy in identifying images and seamless constructions of panoramas. One feature that is yet to be fully utilised is the way an object moves in a video, as most existing research into this area has only looked at doing it with a human as the subject, but it has a potential to work for any object that has points of articulation.Most of the existing research into finding skeletal structures focus only on finding human skeletons and can make the assumption of the shape of a human skeleton beforehand, thus are able to use supervised learning to match the shape to the video that has been provided. The focus of this research is to use a combination of computer vision techniques in order to find the approximate skeletal structure of any object with articulation, with the focus on all types of animals. The reason for using unsupervised learning is that it will allow this to be applied to any species of animal, even though the structures vary vastly so no assumptions of what shapes of body parts from the animal can be made.Applications of such an approach are wide and varied, and cover numerous interdisciplinary domains including: biological image analysis, analysis of various forms of remote sensing data, medical image analysis (for example detecting body parts, and segmenting connected structures within the body (e.g. the heart and lungs)), autonomous vehicle applications (particularly for detecting and modelling articulated vehicles in video footage, which has major applications in e.g. self-driving cars).