We aim at proposing robust and efficient algorithms to detect
meaningful visual events. Robustness implies, in particular, a close control of the number of false alarms made by an algorithm. Since the a contrario statistical approach has proved to match this concern, e.g. to detect geometrical primitives, we extend it to applications where the existing purely analytical framework is not adapted. By combining analytical computations with Monte-Carlo simulations or
statistical learning, we applied a contrario reasoning to problems such as image segmentation into homogeneous regions, which rely on multiple features and on data-driven exploration heuristics whose mathematical properties are difficult to determine.



To satisfy the speed requirement, we also study efficient
architectures. For low level vision, we experimented massive
parallelism and developed a meaningful segments detection algorithm for programmable artificial retina, which operates in real-time. For high level tasks, we propose an agent-based and parallel architecture combining information priorization, parallelism between processing levels and top-down / bottom-up communications to implement anytime algorithms which provide results all along their execution, the most salient first. This architecture is applied to object matching and shows promising results.