Non-invasive multi-view 3D dynamic model extraction
A non-invasive system is presented which is capable of extracting and describing the three-dimensional nature of human gait thereby extending the potential use of gait as a biometric. Of current three-dimensional systems, those using multiple views appear to be the most suitable. Reformulating the three-dimensional anal- ysis algorithm known as Volume Intersection as an evidence gathering process for abstract scene reconstruction provides a new way to overcome concavities and to handle noise and occlusion. After analysis of the standard voxel-based three-dimensional representation, a new data representation called 2.75D is suggested which allows the scene to be analysed at the most appropriate resolution, avoiding further discretisation. With a sequence of three-dimensional frames, another evidence gathering algo- rithm is applied to extract and describe the motion of moving objects. No current techniques have exploited the sequence as a whole during such an operation and in this thesis, a method to incorporate successive frames, and therefore time, as an additional dimension to the extraction process is described. Results on synthetic and real images show that the techniques do indeed process a multi-view image sequence to derive the parameters of interest, thereby provid- ing a suitable basis for future development as a marker-less three-dimensional gait analysis system. In particular, the parameters of a ball moving under the in uence of gravity are extracted with accuracy from a 3D scene. Also, a walking human is extracted and overlaying the result onto the original images conrms that the correct extraction has been made; the result is also supported by medical studies.