Dynamic PoDynamic Point-Cloud Analysis: with int-Cloud Analysis: with applications in Thoracic Surface Reconstruction (TSR) and Autonomous Navigation

点击翻译按钮获取中文摘要

The key aims of the project and novel methodology are summarised below. The research fits in to the computer vision, biomedical engineering and digital-build fields.1. Key Objectives/aims of research: This research will firstly address the problem of fitting objects/surfaces to dynamic point cloud data (such data consists of 3D spatial coordinates of random points on objects over time). This data can be obtained by inexpensive, commercially available time of flight cameras, or from more costly lidar systems. In the case of time of flight cameras, automatically fitting a surface to the thoracic (chest and abdomen) region of a breathing patient, will be one of our applications. A regional analysis of these breathing patterns (and parameters extracted, such as volume and flow) will enable us to automatically classify respiratory disorders (the aim is to look at motor neurone disease and childhood asthma -- with clinical collaborators).In the case of lidar, the point clouds are generally of a stationary object (normally a building) taken with a moving 'camera'. Here we aim to use the motion information to automatically and accurately segment building primitives and to match them to a model, if that exists. This research will tie in with current CUED work in the Cambridge Centre for Digital Build Britain.2. Novel physical sciences/engineering methodology which will be carried out during the course of the project:In order to carry out the research described above we anticipate that we will need to develop a number of novel solutions. Firstly, we will look into novel methods of fitting both geometric primitives (planes etc) and non-uniform surfaces (such as the thoracic region) to dynamic point cloud data. This will require integrating computer vision and computer graphics techniques with statistical optimisation. For the aspects which involve classification we hope to have a new (moving surface over time) dataset, to which we can apply machine learning methods. The conventional CNN (convolutional neural network) methods will require extensions for this data, and the fusion of related data (other known information) could also be important.