Loop-closure detection is crucial for enhancing the robustness of SLAM algorithms in general. For example, after a long travel in unknown terrain, detecting when the robot has returned to a past location makes it possible to increase the accuracy and the consistency of the estimation. Recognizing previously mapped locations can also be relevant for addressing the global localization problem, or even for recovering from a kidnapping (i.e. when the robot has been moved without knowledge of the corresponding displacement). Hence, solving the loop-closure detection problem not only improves SLAM performances, but it affords additional capabilities to mobile robots. The goal of the research effort reported in this thesis is twofold. First, we present a vision-based loopclosure detection algorithm. Our method relies on Bayesian filtering for loop-closure probability computation, with images encoded as sets of local features according to the bags of visual words scheme. When a loop-closure hypothesis receives a high probability, a multiple-view geometry algorithm is employed to discard outliers, by enforcing the existence of a consistent structure between the current image and the loopclosing location. The designed solution is completely incremental, with a linear complexity in the number of places, making it possible to detect loop-closures in real-time conditions. Second, in order to show the benefits of loop-closure detection for mobile robotics, we propose two different applications of our solution to the contexts of topological and metrical SLAM. In the first application, we show how loop-closure detection can be turned into an efficient place recognition module used to build consistent topological maps of the environment : when a new image is acquired, loop-closure detection entails determining if it comes from a new location, or if it pertains to an already existing one, making it possible to update the map in consequence. In the second application, loop-closure detection helps relocating a camera in a metrical SLAM algorithm after a kidnapping : once a mapped part of the environment is recognized, information from the multiple-view geometry algorithm is used to compute a new position and a new orientation for the camera. We demonstrate the quality of our work on indoor, outdoor and mixed (i.e. indoor / outdoor) image sequences acquired using a simple monocular handheld camera in challenging environments, and under strong perceptual aliasing conditions (i.e. when several distinct places look similar).