Abstract
3D Reconstruction using multiple cameras is a technique often employed in the studio production environment. The outdoor sports broadcast environment is also a multicamera environment and this thesis attempts to apply and adapt studio techniques to the unique requirements of this environment, including increased errors in calibration and matting, variations in camera configuration and wide-baseline camera arrangements. This thesis examines 3D reconstruction and activity recognition as applied to the domain of field sports such as football and rugby. Video is captured from a number of cameras arranged around the pitch, and processed using computer vision techniques to generate a 3D scene representation for rendering or analysis. The applications considered in this thesis are free-viewpoint video and the synthetic representation of segments of a match. The thesis presents an analysis of the 3D reconstruction errors in the outdoor sports broadcast environment. Also presented are several Shape-Prom-Silhouette techniques which are shown to provide improved scene reconstruction in this environment. A dual-mode deformable model is presented that refines the reconstruction using stereo information from neighbouring cameras, and simultaneously optimises silhouette extraction in a manner that is more robust to the calibration and reconstruction errors typical of the outdoor sports broadcast scenario. This thesis also presents an action matching technique to extract temporally consistent, synchronised pose information for each player present in the scene. An analysis of matching scores shows that the symmetric Kullback Leibler divergence between shape histograms is a suitable score for measuring the difference between noisy visual hull reconstructions. Players are automatically segmented in 3D and a hidden Markov model is used to match extracted shape histograms against a library of exemplar poses. Various extensions to the scheme are presented and evaluated and it is shown that good results can be achieved using a combination of action matching and key-pose detection.