Real-time Virtual Cinematography for Target Tracking

Soutenance de thèsede Ludovic BURG (VIRTUS) - jeudi 10 novembre 2022 à 09h00 en salle Pétri-Turing

Titre : Real-time Virtual Cinematography for Target Tracking

Cinematic tracking of a moving target in a dynamic 3D environment remains a challenging problem. Proposed systems need to simultaneously ensure low computational cost, a good degree of responsiveness and high cinematic quality despite sudden changes in the environment. In this thesis we address these challenges through two contributions. After presenting the background and domains of the underlying applications considered in this thesis, we perform a thorough literature review of existing methods both in virtual cinematography and drone cinematography. We then present our contributions to solving the real-time cinematic tracking problem. First, we rely on a framing-oriented camera space, the toric space, to track two targets. We extend the approach by proposing a visibility anticipation model and a camera motion model using real-time projection techniques in the toric space. We show the effectiveness of the proposed method through a number of experiments. In the second contribution, we present the design of a new parametric camera representation, the camera animation space. We report the design of a novel trajectory quality metric that we evaluate using GPU-driven ray casting techniques, and couple this representation with a time-horizon planning technique (Model Predictive Control). We designed experimentations to highlight the characteristics of our approach: efficiency, smoothness of trajectories, and visibility anticipation.