Pré-indexation vidéo à l'acquisition, de l'ontologie à l'architecture embarquée

In this thesis we present our research work accomplished in the frame of the MediaMap project. This project has two concurrent objectives concerning the video content : to increase its value for exploration and to perform its early qualification for further processing. Thus, the introduction of a semantic camera of a new generation becomes a fundamental step towards these objectives. The capability to automatically analyze the video content by a video recording device has several advantages, among which the most important are the enhancement of the video content with meta-data and the assistance provided to the amateur in producing video contents of a superior editing quality. Moreover, the functions enabling simultaneously to index and verify the quality of recorded sequences can be directly integrated in the video recording device. In order to achieve this objective, efficient video capture assistance tools need to embedded in the camera, so that the amateur can create high quality content following the guidelines of the integrated scenarios, and qualify the content using the embedded video analysis functions. Even more, the codified and refined scenarios can be further used by profession also in the context search for editing and distribution. The evaluation of the video quality during the recording phase is a challenging problem because it requires a real-time execution of several complex image processing algorithms, in the context of limited hardware resources. In this work we present a range of image processing algorithms that are essential to the semantic video verification and show how their implementation can be optimized in the dynamic execution.